A device early warning method based on intelligent diagnosis, device and medium

Abstract

The application discloses a device early warning method based on intelligent diagnosis, a device and a medium, and belongs to the technical field of device monitoring and early warning. The method comprises the following steps: deploying a temperature sensor network to a device to be detected to obtain original temperature time series data; normalizing the original temperature time series data to generate normalized temperature values; determining a weighted second-order gradient entropy feature based on the normalized temperature values; constructing an end-to-end fault diagnosis model; updating a fault prototype vector weight according to a device working condition change of the device to be detected; determining a confidence score; inputting real-time collected temperature data into the trained end-to-end fault diagnosis model to obtain a fault probability distribution; and early warning when the confidence score exceeds a confidence threshold and the fault probability is higher than a preset probability threshold. The application realizes the technical effects of feature reservation of non-stationary temperature data, collaborative detection of compound faults, multi-scale modeling and optimization of dynamic adaptability through the above method.

Description

Technical Field

[0001] This application relates to the technical field of equipment monitoring and early warning, and in particular to an equipment early warning method, device and medium based on intelligent diagnosis. Background Technology

[0002] In the field of predictive maintenance technology, particularly based on intelligent diagnostics, existing technologies primarily involve deploying high-precision temperature sensor networks to key components of industrial equipment. These sensors collect raw temperature time-series data in real time at a fixed sampling frequency, and the data is transmitted via an industrial IoT gateway to a central processing unit for preprocessing. Existing solutions include standardizing temperature data using conventional normalization methods such as min-max or Z-score to eliminate dimensional differences, and employing differential operations to extract gradient features to identify fault modes. Furthermore, machine learning models such as Long Short-Term Memory (LSTM) networks are widely used in time-series modeling to capture long-term dependencies in temperature data, while attention mechanisms are introduced to extract multi-scale features. Information is then integrated through a feature fusion layer, ultimately outputting a fault probability distribution for maintenance decisions. These technologies have been applied in critical industrial scenarios such as manufacturing and power, aiming to achieve early warning and fault prediction of equipment status through data-driven approaches, thereby reducing downtime and maintenance costs. The entire technology field is rapidly evolving with the development of IoT, AI, and big data processing, providing fundamental support for industrial automation.

[0003] Existing technologies still have significant shortcomings in predictive maintenance of equipment. First, conventional normalization methods ignore the temporal structure of temperature signals, leading to the compression or loss of key features in areas of load fluctuation or abrupt change, and failing to effectively retain abrupt information under non-stationary operating conditions. Second, differential operations are sensitive to high-frequency noise or lag in responding to gradual trends, making it difficult to simultaneously capture significant patterns of complex faults such as bearing wear and cooling shocks within the same framework, reducing the comprehensiveness and reliability of diagnosis. Third, existing models mostly use fixed windows or single LSTM structures, failing to consider temperature change patterns at different scales, resulting in poor identification of early anomalies and small fluctuation precursors, reducing early warning capabilities. Finally, most existing methods use static model parameters and lack dynamic adaptation mechanisms. Once equipment operating conditions change (such as load configuration changes or ambient temperature drift), performance degrades rapidly, making long-term tracking and robust diagnosis impossible. These shortcomings ultimately lead to high false alarm rates, low detection accuracy, and poor long-term stability, limiting the application effectiveness of predictive maintenance in complex industrial environments.

[0004] Therefore, how to achieve feature preservation of non-stationary temperature data, collaborative detection of complex faults, multi-scale modeling, and optimization of dynamic adaptability has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a device early warning method, device, and medium based on intelligent diagnostics to solve the following technical problems: how to achieve feature preservation of non-stationary temperature data, collaborative detection of complex faults, multi-scale modeling, and optimization of dynamic adaptability.

[0006] In a first aspect, embodiments of this application provide a device early warning method based on intelligent diagnosis. The method includes: deploying a high-precision temperature sensor network to key components of the device under test to obtain raw temperature time-series data of the key components, and sending it to a preset central processing unit through a preset industrial IoT gateway; normalizing the raw temperature time-series data through the central processing unit to generate normalized temperature values; determining weighted second-order gradient entropy features based on the normalized temperature values; and constructing an end-to-end fault diagnosis model. The inputs to the end-to-end fault diagnosis model are the normalized temperature values ​​and the weighted second-order gradient entropy features, and the improved long short-term memory network layer of the end-to-end fault diagnosis model captures long-term... The system identifies time-dependent relationships and extracts multi-scale fault features using a multi-head temporal attention algorithm. These features are then processed by an adaptive feature fusion layer to output integrated features. Finally, the integrated features are processed by a fully connected classification layer to output a fault probability distribution. The system updates the fault prototype vector weights based on changes in the operating conditions of the equipment under test. Confidence scores are determined based on the feature deviation, classification uncertainty, and gradient entropy trends in the end-to-end fault diagnosis model. Real-time temperature data is used to generate feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. An early warning is issued when the confidence score exceeds a confidence threshold and the fault probability in the fault probability distribution remains consistently higher than a preset probability threshold.

[0007] In one implementation of this application, the raw temperature time series data is normalized by a central processing unit to generate a normalized temperature value. Specifically, this includes: setting a time window length parameter; dividing consecutively overlapping time windows based on the time window length parameter; wherein each time window includes multiple data points, which are temperature data corresponding to the time series in the raw temperature time series; determining the moving average of the time window based on the arithmetic mean of all data points within the time window; determining the moving standard deviation of the data points within the time window based on the sum of squared deviations between the data points within the window and the moving average; and calculating the temperature data of each time series within the time window based on the moving average and the moving standard deviation to generate a normalized temperature value.

[0008] In one implementation of this application, determining the weighted second-order gradient entropy feature based on normalized temperature values ​​specifically includes: performing a difference operation on the temperature values ​​of any three adjacent time points within the normalized temperature values ​​to determine the second-order difference; wherein, the second-order difference is used to suppress higher-order noise; the calculation of the second-order difference is expressed by the following formula:

[0009]

[0010] in, It is a second-order difference. for The normalized temperature value at time [time]. for The normalized temperature value at time [time]. for The normalized temperature value at any given time;

[0011] Set a fixed-length window. Within the fixed-length time window, divide the second-order difference into multiple gradient value bins according to a preset range to determine the probability distribution of each gradient value bin. Determine the bin weight coefficients based on the mean and maximum gradient of the gradient value bins. Determine the weighted second-order gradient entropy feature based on the probability distribution and the bin weight coefficients.

[0012] In one implementation of this application, the construction of the improved Long Short-Term Memory (LSTM) network layer specifically includes: calculating the average temperature gradient value of the normalized temperature value within a sliding window containing the current time point; multiplying the absolute value of the average temperature gradient value by a preset gating enhancement coefficient to generate a gradient gating term; adding the gradient gating term to the original input term of the forget gate to determine the forget gate input; applying the sigmoid activation function to the enhanced forget gate input to determine the forget gate output; wherein, the forget gate output determines the degree to which the cell state of the previous time step is preserved.

[0013] In one implementation of this application, multi-scale fault features are extracted using a multi-head temporal attention algorithm, specifically including: setting multiple independent attention heads; setting different time window scales for the multiple independent attention heads; wherein, the time window scale is the length of the historical time step range that the attention mechanism focuses on; generating a query vector based on the hidden state output of the long short-term memory network layer at the current time; generating a key vector matrix and a value vector matrix based on the historical hidden state output sequence within the time scale window; calculating the dot product of the query vector and the key vector matrix, dividing by a preset scaling factor, performing masking processing, and applying the softmax function to obtain the weight distribution; performing a weighted summation of the value vector matrix based on the weight distribution to determine the output vector of the independent attention head at the current time window scale; and concatenating the output vectors of the multiple independent attention heads to output the multi-scale fault features.

[0014] In one implementation of this application, the multi-scale fault features are output as integrated features by an adaptive feature fusion layer. Specifically, this includes: concatenating the current hidden state output of the Long Short-Term Memory network with the feature vector output by the multi-head temporal attention algorithm to obtain a concatenated feature vector; processing the concatenated feature vector based on the sigmoid activation function to generate a fusion gate vector; determining the weighted hidden state features based on the hidden state output of the Long Short-Term Memory network layer and the fusion gate vector; determining the weighted attention features based on the feature vector output by the multi-head temporal attention algorithm and the fusion gate vector; and generating integrated features based on the weighted hidden state features and the weighted attention features.

[0015] In one implementation of this application, the fault prototype vector weights are updated according to changes in the operating conditions of the device under test, specifically including:

[0016] based on Determine the inter-class distance of the constraint to determine the prototype contrast loss; where, For class prototype; For prototype comparison loss, Number of fault categories For the first Class sample feature mean, For category Sample feature output, For boundary margin, For category Sample feature output, For category Sample feature output, For Euclidean distance For category The number of samples, Let i be the set of samples of class i. For sample index, For feature extraction networks, For the nth sample, This represents the function that takes the maximum value.

[0017] A preset cross-entropy loss function is introduced; the total loss function is determined based on the prototype contrast loss and the cross-entropy loss function; the new batch data and historical batch data of the device to be tested are determined; when the batch difference between the new batch feature distribution of the new batch data and the historical feature distribution of the historical batch data is greater than the preset difference threshold, the update rate is determined according to the value of the batch difference; the fault prototype vector weights are updated based on the update rate.

[0018] In one implementation of this application, the confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trend in the end-to-end fault diagnosis model. Specifically, this includes: measuring the Euclidean distance between the current fused feature vector and the feature prototype vector of the end-to-end fault diagnosis model, and dividing it by the maximum norm of all category feature prototype vectors to determine the feature deviation; calculating the fault probability distribution output by the end-to-end fault diagnosis model to determine the information entropy, thereby determining the classification uncertainty; wherein, a higher entropy value for classification uncertainty indicates greater uncertainty; determining the gradient entropy change based on the absolute value of the difference between the weighted second-order gradient entropy feature value at the current time and the second-order gradient entropy feature value at the previous time; multiplying the feature deviation, classification uncertainty, and gradient entropy change by preset weight coefficients and then summing them to determine a weighted comprehensive value; and inputting the weighted comprehensive value into a sigmoid function for nonlinear mapping to determine the confidence score.

[0019] Secondly, embodiments of this application also provide a device early warning device based on intelligent diagnosis. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to: deploy a high-precision temperature sensor network to key components of the device under test to obtain raw temperature time-series data of the key components and send it to a preset central processing unit via a preset industrial IoT gateway; normalize the raw temperature time-series data through the central processing unit to generate a normalized temperature value; determine a weighted second-order gradient entropy feature based on the normalized temperature value; and construct an end-to-end fault diagnosis model. The input to the end-to-end fault diagnosis model is the normalized temperature... The system captures long-term dependencies using an improved long short-term memory network layer in the end-to-end fault diagnosis model, based on the degree value and weighted second-order gradient entropy features. Multi-scale fault features are extracted using a multi-head temporal attention algorithm. These multi-scale features are then processed by an adaptive feature fusion layer to output integrated features. Finally, the integrated features are processed by a fully connected classification layer to output a fault probability distribution. The system updates the fault prototype vector weights based on changes in the equipment's operating conditions. A confidence score is determined based on feature deviation, classification uncertainty, and gradient entropy trends in the end-to-end fault diagnosis model. Real-time temperature data is used to generate feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. An early warning is issued when the confidence score exceeds a confidence threshold and the fault probability in the fault probability distribution remains consistently higher than a preset probability threshold.

[0020] Thirdly, this application also provides a non-volatile deterministic machine storage medium for device early warning based on intelligent diagnosis, storing deterministic machine executable instructions. These instructions are configured to: deploy a high-precision temperature sensor network to key components of the device under test to obtain raw temperature time-series data of the key components, and send this data to a preset central processing unit via a preset industrial IoT gateway; normalize the raw temperature time-series data through the central processing unit to generate a normalized temperature value; determine a weighted second-order gradient entropy feature based on the normalized temperature value; and construct an end-to-end fault diagnosis model. The inputs to the end-to-end fault diagnosis model are the normalized temperature value and the weighted second-order gradient entropy feature. An improved Long Short-Term Memory (LSTM) network layer captures long-term dependencies. Multi-scale fault features are extracted using a multi-head temporal attention algorithm. These multi-scale fault features are then processed by an adaptive feature fusion layer to output integrated features. Finally, the integrated features are processed by a fully connected classification layer to output a fault probability distribution. The fault prototype vector weights are updated based on changes in the operating conditions of the device under test. The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trends in the end-to-end fault diagnosis model. Real-time temperature data is used to generate feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. An early warning is issued when the confidence score exceeds a confidence threshold and the fault probability in the fault probability distribution remains higher than a preset probability threshold.

[0021] This application provides a device early warning method, device, and medium based on intelligent diagnostics, which at least includes the following technical effects:

[0022] In terms of data preprocessing, the piecewise adaptive normalization method effectively improves the feature retention capability of non-stationary temperature data. By dividing the data into dynamic time windows to calculate the moving mean and standard deviation, local normalization is performed at each time point. This method solves the problem of key feature loss caused by conventional min-max or Z-score normalization in regions of load fluctuation or abrupt change.

[0023] In the feature extraction stage, the weighted second-order gradient entropy feature significantly enhances the detection sensitivity and noise resistance of composite faults. By calculating the second-order difference to suppress high-frequency noise and combining it with gradient distribution weighted binning statistics, this method amplifies the entropy changes of slowly varying faults (such as bearing wear) and sudden faults (such as cooling shocks). This method exhibits a continuously enhanced response to slowly varying faults and generates pure pulse peaks for instantaneous faults, solving the problems of response lag or noise sensitivity in traditional differential operations, thereby improving the comprehensiveness and stability of diagnosis.

[0024] In terms of model architecture design, the end-to-end fault diagnosis model significantly improves multi-scale feature fusion capability and discrimination accuracy by improving the long short-term memory network, multi-head temporal attention mechanism, and adaptive feature fusion layer. The improved LSTM forgetting gate combined with temperature gradient gating enhances the long-term memory capability for slowly varying faults; the multi-head attention mechanism extracts fault features from different time windows in parallel, overcoming the limitations of single-scale modeling; and the adaptive feature fusion layer dynamically weights features based on temperature fluctuation variance, improving the model's adaptability to complex fault modes.

[0025] To address the issue of equipment state drift, the online dynamic weight update mechanism significantly enhances the long-term robustness of the model. Through an incremental prototype adjustment strategy, the prototype vector weights are adaptively updated when the feature distribution of the new batch differs significantly from the historical distribution, thus resolving the performance degradation problem of the static model under changing operating conditions (such as load configuration changes or ambient temperature drift).

[0026] In the fault early warning stage, the multi-index confidence quantification method significantly improves the accuracy and reliability of alarms. By jointly evaluating feature deviation, classification uncertainty, and gradient entropy change trends, a comprehensive confidence score is generated, solving the problem of high false alarm rate in traditional threshold methods. This method ensures, to a certain extent, that when the confidence level exceeds the threshold and the fault probability remains high, the triggered alarm has high credibility, thereby optimizing the practicality of maintenance decisions. Attached Figure Description

[0027] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0028] Figure 1 A flowchart of a device early warning method based on intelligent diagnosis provided in this application embodiment;

[0029] Figure 2 This is a schematic diagram of the original temperature signal;

[0030] Figure 3 A comparison chart showing the effects of different normalization methods;

[0031] Figure 4 for Figure 3 Enlarged view of the medium-Gaussian fault area;

[0032] Figure 5 A comparison chart of different feature extraction methods and the feature extraction method of this application;

[0033] Figure 6 A comparison chart of ablation experiment performance for different model architectures;

[0034] Figure 7A comparison chart showing the performance changes of different prediction methods over long-term operation;

[0035] Figure 8 This is a schematic diagram of the internal structure of a device early warning device based on intelligent diagnosis, provided as an embodiment of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] This application provides a device early warning method, device, and medium based on intelligent diagnostics to solve the following technical problems: how to achieve feature preservation of non-stationary temperature data, collaborative detection of complex faults, multi-scale modeling, and optimization of dynamic adaptability.

[0038] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0039] Figure 1 This document provides a flowchart of a device early warning process based on intelligent diagnostics, as illustrated in an embodiment of this application. Figure 1 As shown in the figure, the device early warning method based on intelligent diagnosis provided in this application embodiment specifically includes the following steps:

[0040] Step 1: Deploy a high-precision temperature sensor network to the key components of the device under test to obtain the raw temperature time-series data of the key components, and send it to the preset central processing unit through the preset industrial IoT gateway.

[0041] A "high-precision temperature sensor network" refers to a group of temperature sensing devices distributed on critical operating components of equipment (such as bearings, motors, or cooling systems). These sensors have high resolution (e.g., ±0.1°C accuracy) and a fixed sampling frequency (e.g., 1Hz) for continuous monitoring of temperature changes. "Critical components" refer to the core components of the equipment that are prone to failure or affect overall performance, ensuring that data acquisition covers the entire operating range (including startup, steady-state operation, load fluctuations, and shutdown).

[0042] An "Industrial IoT gateway" is a data relay device responsible for transmitting raw temperature time-series data collected by sensors to a central processing unit in real time, ensuring data synchronization and integrity. A "preset central processing unit" refers to a remote or local computing server used for subsequent data processing.

[0043] In a specific example, equipment A required predictive maintenance. The maintenance team deployed three high-precision temperature sensors (sampling frequency 1Hz) at the agitator shaft seal of the reactor to collect real-time temperature change signals during equipment operation. These sensors synchronously transmitted the data to the central processing unit via the factory's industrial IoT gateway. The data acquisition process covered the entire operating range of the equipment, including the reactor startup phase (temperature rising from room temperature to 150°C), the steady-state operation phase (temperature fluctuation ±2°C), the load surge phase (simulating sudden temperature changes caused by feed changes), and the shutdown cooling phase. Through this process, a time-series temperature dataset containing both normal conditions and potential seal failure modes was obtained.

[0044] Step 2: Normalize the raw temperature time series data through the central processing unit to generate normalized temperature values.

[0045] Step 2.1: Set the time window length parameter and divide the time window into consecutive overlapping time windows based on the time window length parameter; wherein, the time window includes multiple data points, and the data points are the temperature data corresponding to the time series in the original temperature time series.

[0046] First, determine the time window length τ and half-window width w to divide multiple consecutive overlapping time windows, and then follow the formula... Determine the time window number k to which the current time point t belongs.

[0047] Step 2.2: Determine the moving average of the time window based on the arithmetic mean of all data points within the time window.

[0048] Furthermore, within the time window number k, the formula is used... Determine the moving average; where, It is the moving average. For time indexing, for Original temperature value at any time This represents the total number of data points within the window. .

[0049] Step 2.3: Determine the sliding standard deviation of the data points within the time window by the sum of squared deviations between the data points within the window and the moving mean.

[0050] Furthermore, through the formula Determine the sliding standard deviation; where, For the sliding standard deviation, The original temperature value at time t.

[0051] Step 2.4: Calculate the temperature data for each time series within the time window based on the moving mean and moving standard deviation to generate normalized temperature values.

[0052] Furthermore, according to the formula Generate normalized temperature values; where γ k β is the scaling factor. k ε is the offset factor, and ε is the smoothing factor. This is the normalized temperature value.

[0053] In a specific example, for temperature monitoring data of device A (including non-stationary characteristics caused by load fluctuations), the central processing unit performs piecewise adaptive normalization processing. The time window length τ = 60 seconds (window length) and the half-window width w = 10 seconds (window radius) are set. For the data point at t = 300 seconds, the moving mean (average of all temperature values ​​within 300 ± 10 seconds) and moving standard deviation within its corresponding window (k = 300 / 60 = 5) are calculated.

[0054] like Figures 2 to 4 As shown, to verify the feature preservation capability of the piecewise adaptive normalization method under complex working conditions, the method of this invention is compared with the minimum-maximum normalization method (…). Figure 3 Min-Max normalization), Z-score normalization ( Figure 3 We compared the three normalization methods (Z-Score normalization) and analyzed their feature preservation effects on non-stationary temperature signals (including load fluctuations, instantaneous spikes, Gaussian faults, and exponentially rising faults).

[0055] like Figure 2 As shown, the original temperature signal display device experienced multiple load changes and three typical fault events.

[0056] Traditional normalization methods (min-maximum, Z-score) exhibit characteristic smoothing in the fault region, such as... Figure 4 As shown, especially in the Gaussian fault region (enlarged view of 2000-2100 points), key abrupt change features are significantly weakened. The method of this invention fully preserves the detailed features of all fault events: the steep rising edge of the instantaneous peak, the complete waveform of the Gaussian fault, and the gradual trend of exponential increase are all clearly visible.

[0057] Experimental results demonstrate that piecewise adaptive normalization, through dynamic time window statistics, effectively solves the feature loss problem in non-stationary signal processing.

[0058] Step 3: Determine the weighted second-order gradient entropy characteristics based on the normalized temperature value.

[0059] Step 3.1: Perform a difference operation on the temperature values ​​of any three adjacent time points in the normalized temperature value to determine the second-order difference; the second-order difference is used to suppress higher-order noise.

[0060] The second-order difference is calculated using the following formula:

[0061]

[0062] in, It is a second-order difference. for The normalized temperature value at time [time]. for The normalized temperature value at time [time]. for The normalized temperature value at any given time;

[0063] Step 3.2: Set a fixed-length window. Within the fixed-length time window, divide the second-order difference into multiple gradient value bins according to a preset range to determine the probability distribution of each gradient value bin.

[0064] Furthermore, a gradient value bin set B is preset, and the probability of the second difference value falling into each bin b is statistically calculated within a time window of length L based on the gradient value bin set B. To construct gradient bins; where, Before and after time Gradient values ​​fall into the bin within the time window The probability, For counting functions, The time window length, This is a binning index.

[0065] Step 3.3: Determine the bin weight coefficients based on the mean and maximum gradient of the bins.

[0066] The bin weighting coefficient is expressed by the following formula:

[0067]

[0068] in, For the first Mean gradient within each bin It represents the maximum absolute value of the second-order gradient within the current time window.

[0069] Step 3.4: Determine the weighted second-order gradient entropy features based on the probability distribution and bin weight coefficients.

[0070] Furthermore, through the formula Output the weighted second-order gradient entropy features; where, It is the weighted second-order gradient entropy.

[0071] In a specific example, for temperature monitoring data of device A, the weighted second-order gradient entropy feature is first extracted. First, the second difference of the normalized temperature sequence is calculated (value at time t = (temperature at time t+1 - 2 × temperature at time t + temperature at time t-1)) to suppress high-frequency noise interference. A gradient bin set B is preset (-0.5~0.5℃ / second, one bin per 0.1℃ / second), and the probability of each bin is statistically analyzed within a time window of L = 30 seconds (e.g., the proportion of samples falling within the -0.1~0℃ / second range). Finally, the entropy feature is output by combining the bin weights (positively correlated with the absolute value of the average gradient within the bin), enhancing the sensitivity to bearing wear (gradual failure) and cooling failure (sudden failure).

[0072] like Figure 5 As shown, the comparison of feature extraction methods in terms of fault sensitivity is analyzed to verify the sensitivity of weighted second-order gradient entropy for detecting composite faults. By comparing the response characteristics of the feature extraction method of this invention with first-order and second-order gradient features in the composite scenarios of bearing wear (gradual fault) and cooling system impact (instantaneous fault), it is shown that the first-order gradient feature is sensitive to cooling impact but suffers from severe noise interference, with almost no response in the bearing wear region. Although the second-order gradient feature suppresses noise, its response to gradual faults is lagging and insufficient in amplitude. The method of this invention exhibits optimal characteristics in both types of faults, showing a gradually increasing response to bearing wear (yellow area after 100 seconds), producing significant and pure pulse peaks for three cooling impacts (red area), maintaining a stable baseline in non-fault areas, and demonstrating significant noise suppression. This feature achieves collaborative detection of gradual / abrupt faults through the fusion design of entropy weighting and gradient distribution statistics.

[0073] Step 4: Construct an end-to-end fault diagnosis model. The input of the end-to-end fault diagnosis model is the normalized temperature value and the weighted second-order gradient entropy feature. The improved long short-term memory network layer of the end-to-end fault diagnosis model captures long-term dependencies. Multi-scale fault features are extracted by a multi-head temporal attention algorithm. The multi-scale fault features are output as integrated features by an adaptive feature fusion layer. The integrated features are output as fault probability distribution by a fully connected classification layer.

[0074] Step A1: Calculate the average temperature gradient of the normalized temperature value within a sliding window that includes the current time point.

[0075] According to the formula Determine the temperature gradient.

[0076] in, for Normalized temperature value at any time for Normalized temperature value at any time For the current time step, For sliding index variables, for Average temperature gradient inside the window for Normalized temperature value at any time.

[0077] Step A2: Multiply the absolute value of the average temperature gradient by a preset gating enhancement coefficient to generate a gradient gating term.

[0078] Step A3: Add the gradient gating term to the original input term of the forget gate to determine the forget gate input.

[0079] Step A4: Apply the sigmoid activation function to the enhanced forget gate input to determine the forget gate output; whereby the forget gate output determines the degree to which the cell state of the previous moment is preserved.

[0080] Steps A2-A4 are expressed by the following formulas:

[0081]

[0082] in, , This is the normalized temperature value. The vector is a concatenated weighted second-order gradient entropy vector, where β is the gated enhancement coefficient. Forget the gate output, Output for the forget gate. for Normalized temperature value at any time For the sigmoid function, for Normalized temperature value at any time This is the hidden state from the previous moment.

[0083] Furthermore, multi-scale fault features are extracted using a multi-head temporal attention algorithm, including:

[0084] B1. Set up multiple independent attention heads.

[0085] Set up multiple independent attention heads; among them, the multiple independent attention heads include short window independent attention head, medium window independent attention head, and long window independent attention head.

[0086] B2. Set different time window scales for multiple independent attention heads; where the time window scale is the length of the historical time step range that the attention mechanism focuses on.

[0087] B3. Generate a query vector based on the hidden state output of the Long Short-Term Memory network layer at the current moment.

[0088] B4. Generate a key vector matrix and a value vector matrix based on the historical hidden state output sequence within the time scale window.

[0089] B5. Calculate the dot product of the query vector and the key vector matrix, divide by the preset scaling factor, perform masking, and apply the softmax function to obtain the weight distribution.

[0090] B6. Based on the weight distribution, perform a weighted summation of the value vector matrix to determine the output vector of the independent attention head at the current time window scale.

[0091] B2 to B6 are represented by the following formulas:

[0092]

[0093] in, For the first The output of each attention head, the attention function is the standard dot product attention function, and , For querying the matrix, For the first The query weight matrix for each attention head. The key matrix, For the first The key weight matrix of each attention head. For value matrices, For the first The value weight matrix of each attention head, For the softmax function, Scaling factor This is the transpose of the key matrix. It is a lower triangular mask matrix. This is element-wise multiplication.

[0094] B7. Concatenate the output vectors of multiple independent attention heads to output multi-scale fault features.

[0095] The attention features output by multiple independent attention heads are calculated using the formula splicing; among which, For the output of multi-head temporal attention, This is a feature vector concatenation operation. The number of attention heads.

[0096] Furthermore, the multi-scale fault features are integrated into an adaptive feature fusion layer, which includes:

[0097] C1. Concatenate the current hidden state output of the Long Short-Term Memory network with the feature vector output by the multi-head temporal attention algorithm to obtain the concatenated feature vector.

[0098]

[0099] in, Based on the weight matrix, This is the weighting adjustment factor. Temperature variance within the current time window

[0100] C2. Process the concatenated feature vectors based on the sigmoid activation function to generate a fused gated value vector.

[0101] Steps C1 and C2 are represented by the following formulas:

[0102]

[0103] in, To fuse the gated value vector, This represents the original hidden state of the LSTM. The bias vector for fusion gating. This is the sigmoid function.

[0104] C3. Based on the hidden state output of the long short-term memory network layer and the fusion gate value vector, determine the weighted hidden state features.

[0105] C4. Based on the feature vector output by the multi-head temporal attention algorithm and the fusion gate value vector, the weighted attention features are determined.

[0106] Steps C3 and C4 are represented by the following formulas:

[0107]

[0108] in, The fused feature vector These are the weighted hidden state features. These are the weighted attention features.

[0109] C5. Generate integrated features based on weighted latent state features and weighted attention features.

[0110] The fused features are input into a fully connected layer, activated by a softmax function, and then output. ;in, This represents the failure probability distribution.

[0111] In a specific case, device A needs to build an end-to-end fault diagnosis model. First, improve the LSTM layer: calculate the sliding window (…). The average temperature gradient over a 20-second period (the average rate of temperature change between the current moment and the previous 20 seconds) is injected into the forget gate. When the equipment experiences gradual bearing wear, this mechanism allows the forget gate to retain a long-term memory of the temperature increase trend, reducing the false negative rate by 40%. Secondly, a multi-head attention mechanism is implemented: a short window head (10 steps) captures instantaneous impacts on the cooling system, a medium window head (30 steps) identifies abnormal load fluctuations, and a long window head (60 steps) tracks gear wear trends. Each head calculates feature weights using standard dot product attention, extracting multi-scale information in parallel. Finally, adaptive feature fusion is performed: based on the temperature variance of the current time window (e.g., when the variance > 2℃), the gating network dynamically increases the attention feature weights (e.g., from 0.3 to 0.7) to address the insufficient adaptability of fixed fusion to sudden faults. The fused features are then output as fault probabilities (e.g., 85% probability for "bearing wear") via a fully connected layer.

[0112] like Figure 6 As shown, ablation experiments were conducted to compare the model architectures and verify the necessity of key components in the end-to-end diagnostic model, using radar. Figure 6 The model is compared across multiple metrics (accuracy, recall, F1 score, area under the curve, early detection rate, and false positive rate) with a simplified version that removes multi-head attention and adaptive fusion. Experimental results show that the full model achieves the largest envelope across all metrics, with significant advantages in early detection rate and false positive rate. Removing multi-head attention leads to a decrease in multi-scale feature extraction capability, with the area under the curve shrinking most noticeably. Removing the adaptive fusion layer significantly reduces recall (increases missed detections), demonstrating the crucial role of dynamic feature fusion in adaptability to various operating conditions. The traditional long short-term memory network baseline model performs the weakest across all dimensions, validating the effectiveness of gradient gating for remembering slowly changing faults.

[0113] Step 5: Update the fault prototype vector weights based on changes in the operating conditions of the equipment under test.

[0114] Step 5.1, based on Determine the inter-class distance of the constraint to determine the prototype contrast loss; where, For class prototype; For prototype comparison loss, Number of fault categories For the first Class sample feature mean, For category Sample feature output, For boundary margin, For category Sample feature output, For category Sample feature output, For Euclidean distance For category The number of samples, Let i be the set of samples of class i. For sample index, For feature extraction networks, For the nth sample, This represents the function that takes the maximum value.

[0115] Step 5.2: Introduce the preset cross-entropy loss function.

[0116] Introducing a pre-defined cross-entropy loss function .

[0117] Step 5.3: Determine the total loss function based on the prototype contrast loss and cross-entropy loss function.

[0118]

[0119] in, For the total loss function, Let cross-entropy be the loss function. It is a balancing factor.

[0120] Step 5.4: Determine the new batch data and historical batch data of the equipment to be tested.

[0121] Step 5.5: When the batch difference between the new batch feature distribution of the new batch data and the historical feature distribution of the historical batch data is greater than the preset difference threshold, the update rate is determined based on the value of the batch difference.

[0122]

[0123]

[0124] in, For the updated number class prototype vector, To achieve adaptive update rate, To achieve adaptive update rate, Category in the current batch The sample, For batch medium sample Input features, For sample index, The characteristic covariance matrix, It is the Frobenius norm. Based on the learning rate, The drift sensitivity coefficient, It is an exponential function with base to natural numbers. Categories in historical datasets All samples.

[0125] Step 5.6: Update the fault prototype vector weights based on the update rate.

[0126] In a specific example, device A experiences temperature characteristic drift due to a load configuration upgrade. First, the prototype contrast loss is calculated: the mean feature value is extracted from the historical dataset (containing three types of samples: normal, bearing wear, and cooling failure) as the class prototype (e.g., the bearing wear prototype). During training, constrain the features of similar samples and The distance between the target and non-target classes is less than the distance between the target and non-target classes (margin m=0.5). The total loss function is combined with cross-entropy loss (weight λ=0.8) to improve the clarity of the classification boundary. When the difference between the feature distribution of the new batch of data and the historical distribution exceeds a threshold, an incremental update is triggered: the adaptive update rate η is calculated based on the difference (base learning rate). =0.1, drift sensitivity coefficient =0.3), updated prototype vector Integrate new sample features (e.g., weight η=0.15).

[0127] Step 6: Determine the confidence score based on the feature deviation, classification uncertainty, and gradient entropy change trend in the end-to-end fault diagnosis model.

[0128] Step 6.1: Measure the Euclidean distance between the current fused feature vector and the feature prototype vector of the end-to-end fault diagnosis model, and divide it by the maximum norm of all category feature prototype vectors to determine the feature deviation.

[0129] Feature deviation from express;

[0130] in, The fused feature vector For predicting categories The prototype, It is the maximum value among all class prototype norms. Let be the mean of the features of the i-th class of samples.

[0131] Step 6.2: Calculate the fault probability distribution output by the end-to-end fault diagnosis model to determine the information entropy and thus the classification uncertainty; where the higher the entropy value of the classification uncertainty, the greater the uncertainty.

[0132]

[0133] in, Due to classification uncertainty, Output probabilities for the model.

[0134] Step 6.3: Determine the change in gradient entropy based on the absolute value of the difference between the weighted second-order gradient entropy eigenvalue at the current moment and the second-order gradient entropy eigenvalue at the previous moment.

[0135]

[0136] in, for Time-weighted second-order gradient entropy, for The weighted second-order gradient entropy at each time step.

[0137] Step 6.4: Multiply the feature deviation, classification uncertainty, and gradient entropy change by preset weight coefficients and then sum them to determine the weighted composite value.

[0138] Step 6.5: Input the weighted composite value into the sigmoid function for nonlinear mapping to determine the confidence score.

[0139] Steps 6.4 and 6.5 are expressed by the following formulas:

[0140]

[0141] in, The fault confidence score. The fault confidence score. For classifying uncertainty weights, The weights represent the gradient entropy changes.

[0142] In a specific example, during online diagnostics of device A, the confidence level is calculated in real time. This involves obtaining the feature deviation (distance ratio between the current feature and the "sealing failure" class prototype = 0.9), classification uncertainty (entropy value of the output probability distribution = 0.2), and gradient entropy change (…). =0.3). Generate a confidence score according to the preset weights (w1=0.6, w2=0.3, w3=0.1): sig(0.6×0.9+0.3×0.2+0.1×0.3)=0.85.

[0143] Step 7: Generate feature vectors from the real-time collected temperature data and input them into the trained end-to-end fault diagnosis model to obtain the fault probability distribution.

[0144] In a specific example, during real-time monitoring of device A, the central processing unit performs normalization (window length τ = 30 seconds) and weighted second-order gradient entropy extraction (bin set B = 5 intervals) on the temperature data collected every second. These two processes are concatenated into a feature vector (dimension 64), which is then input into the end-to-end model trained in step 4. After LSTM temporal modeling and multi-head attention fusion, a multi-fault probability distribution is output (e.g., "bearing wear" probability 75%, "cooling failure" probability 10%, "normal" probability 15%).

[0145] Step 8: When the confidence score exceeds the confidence threshold and the failure probability in the failure probability distribution is consistently higher than the preset probability threshold, an early warning is issued.

[0146] In a specific case, during the diagnosis of device A, the confidence score was 0.78 (threshold 0.7), and the probability of "compressor failure" was higher than 85% for 10 consecutive minutes. This triggered a level-two warning: an alarm work order was sent to the maintenance terminal, and the backup cooling unit was activated.

[0147] like Figure 7 As shown, a comparison of dynamic adaptability during long-term operation is conducted to verify the adaptability of the dynamic weight update mechanism to equipment state drift. Within a 100-day operating cycle, the accuracy stability of this invention, a fixed model, a conventional long short-term memory network, and traditional methods is compared under four equipment state changes (load configuration change / cooling system upgrade / ambient temperature change / equipment aging). The fixed model exhibits a step-wise performance degradation after each state change, with the degradation intensifying during the equipment aging phase. The conventional long short-term memory network is sensitive to ambient temperature changes, exhibiting the largest accuracy fluctuation. The method of this invention quickly recovers high accuracy after all four state changes, indicating that the dynamic prototype adjustment mechanism, through the perception of feature distribution differences, achieves incremental optimization of model parameters, enabling the diagnostic system to continuously evolve.

[0148] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a device early warning device based on intelligent diagnosis, the structure of which is as follows: Figure 2 As shown.

[0149] Figure 8 This is a schematic diagram of the internal structure of a device early warning system based on intelligent diagnostics, provided as an embodiment of this application. Figure 8 As shown, the device includes:

[0150] At least one processor 801;

[0151] And a memory 802 that is communicatively connected to at least one processor;

[0152] The memory 802 stores instructions executable by at least one processor, which are executed by at least one processor 801 to enable at least one processor 801 to:

[0153] A high-precision temperature sensor network is deployed to the key components of the device under test to obtain raw temperature time-series data of the key components, and sent to a preset central processing unit via a preset industrial IoT gateway. The central processing unit normalizes the raw temperature time-series data to generate normalized temperature values. Weighted second-order gradient entropy features are determined based on the normalized temperature values. An end-to-end fault diagnosis model is constructed. The inputs to the end-to-end fault diagnosis model are the normalized temperature values ​​and the weighted second-order gradient entropy features. Long-term dependencies are captured through an improved long short-term memory network layer of the end-to-end fault diagnosis model, and multi-head temporal attention algorithms are used to improve the accuracy of the data. Multi-scale fault features are extracted, and the multi-scale fault features are output as integrated features through an adaptive feature fusion layer. The integrated features are then output as a fault probability distribution through a fully connected classification layer. The weights of the fault prototype vector are updated according to the changes in the operating conditions of the equipment under test. The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trends in the end-to-end fault diagnosis model. The real-time collected temperature data is used to generate feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. When the confidence score exceeds the confidence threshold and the fault probability in the fault probability distribution remains higher than the preset probability threshold, an early warning is issued.

[0154] Some embodiments of this application provide corresponding to Figure 1 A non-volatile deterministic machine storage medium for device early warning based on intelligent diagnostics, storing deterministic machine executable instructions, wherein the deterministic machine executable instructions are configured as follows:

[0155] A high-precision temperature sensor network is deployed to the key components of the device under test to obtain raw temperature time-series data of the key components, and sent to a preset central processing unit via a preset industrial IoT gateway. The central processing unit normalizes the raw temperature time-series data to generate normalized temperature values. Weighted second-order gradient entropy features are determined based on the normalized temperature values. An end-to-end fault diagnosis model is constructed. The inputs to the end-to-end fault diagnosis model are the normalized temperature values ​​and the weighted second-order gradient entropy features. Long-term dependencies are captured through an improved long short-term memory network layer of the end-to-end fault diagnosis model, and multi-head temporal attention algorithms are used to improve the accuracy of the data. Multi-scale fault features are extracted, and the multi-scale fault features are output as integrated features through an adaptive feature fusion layer. The integrated features are then output as a fault probability distribution through a fully connected classification layer. The weights of the fault prototype vector are updated according to the changes in the operating conditions of the equipment under test. The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trends in the end-to-end fault diagnosis model. The real-time collected temperature data is used to generate feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. When the confidence score exceeds the confidence threshold and the fault probability in the fault probability distribution remains higher than the preset probability threshold, an early warning is issued.

[0156] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0157] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0158] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or deterministic program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a deterministic program product implemented on one or more deterministic storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing deterministic program code.

[0159] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and deterministic machine program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by deterministic machine program instructions. These deterministic machine program instructions can be provided to a processor of a general-purpose deterministic machine, a special-purpose deterministic machine, an embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the deterministic machine or other programmable data processing apparatus, produce implementations of the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0160] These deterministic machine program instructions may also be stored in a deterministic machine-readable memory that can direct a deterministic machine or other programmable data processing device to operate in a particular manner, such that the instructions stored in the deterministic machine-readable memory produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0161] These deterministic machine program instructions can also be loaded onto a deterministic machine or other programmable data processing device, causing a series of operational steps to be performed on the deterministic machine or other programmable device to produce a process implemented in the deterministic machine, thereby providing instructions that execute on the deterministic machine or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0162] In a typical configuration, a device is defined as having one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0163] Memory may include forms such as non-persistent storage in machine-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a machine-readable medium.

[0164] Deterministic machine-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be deterministic machine-readable instructions, data structures, program modules, or other data. Examples of deterministic machine-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a deterministic device. As defined herein, deterministic machine-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0165] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0166] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A device early warning method based on intelligent diagnosis, characterized in that, The method comprises: deploying a high-precision temperature sensor network to the key components of the equipment to be detected to obtain original temperature time series data of the key components and send the data to a preset central processing unit through a preset industrial Internet of Things gateway; normalizing the original temperature time series data through the central processing unit to generate normalized temperature values; determining a weighted second-order gradient entropy feature based on the normalized temperature values; constructing an end-to-end fault diagnosis model; wherein the input of the end-to-end fault diagnosis model is the normalized temperature values and the weighted second-order gradient entropy feature, the long-term dependence relationship is captured through an improved long short-term memory network layer of the end-to-end fault diagnosis model, and multi-scale fault features are extracted through a multi-head time sequence attention algorithm, the multi-scale fault features are output through a self-adaptive feature fusion layer to integrate features, and the integrated features are output through a fully connected classification layer to obtain a fault probability distribution; updating the fault prototype vector weight according to the equipment working condition change of the equipment to be detected; determining a confidence score according to the feature deviation degree, classification uncertainty and gradient entropy change trend in the end-to-end fault diagnosis model; generating feature vectors from the real-time collected temperature data in sequence and inputting the feature vectors into the trained end-to-end fault diagnosis model to obtain a fault probability distribution; issuing a warning when the confidence score exceeds a confidence threshold and the fault probability in the fault probability distribution continuously exceeds a preset probability threshold; The construction of the improved long short-term memory network layer specifically comprises: calculating the average temperature gradient value of the normalized temperature values in a sliding window containing the current time point; multiplying the absolute value of the average temperature gradient value by a preset gate enhancement coefficient to generate a gradient gate term; adding the gradient gate term to the original input term of the forget gate to determine the forget gate input; applying a sigmoid activation function to the enhanced forget gate input to determine the forget gate output; wherein the forget gate output determines the degree to which the cell state of the previous time is retained.

2. The device early warning method based on intelligent diagnosis according to claim 1, characterized in that, The normalization of the original temperature time series data through the central processing unit to generate normalized temperature values specifically comprises: setting a time window length parameter and dividing continuous overlapping time windows based on the time window length parameter; wherein the time window includes a plurality of data points, and the data points are temperature data corresponding to the time series in the original temperature time series; determining the sliding mean of the time window based on the arithmetic mean of all data points in the time window; determining the sliding standard deviation of the data points in the time window through the sum of squares of deviations of the data points in the window from the sliding mean; calculating the temperature data of each time series in the time window based on the sliding mean and the sliding standard deviation to generate normalized temperature values.

3. The device early warning method based on intelligent diagnosis according to claim 2, characterized in that, The determination of the weighted second-order gradient entropy feature based on the normalized temperature values specifically comprises: performing a difference operation on the temperature values of any three adjacent time points in the normalized temperature values to determine a second-order difference; wherein the second-order difference is used to suppress high-order noise; the calculation of the second-order difference is represented by the following formula: wherein, is a second order difference, is a normalized temperature value at time instant, is a normalized temperature value at time instant, is a normalized temperature value at time instant; A fixed-length window is set, and within the fixed-length time window, the second-order difference is divided into multiple gradient value bins according to a preset range to determine the probability distribution of each gradient value bin. The bin weight coefficients are determined based on the mean and maximum gradient of the bins according to the gradient values. The weighted second-order gradient entropy feature is determined based on the probability distribution and the bin weight coefficients.

4. The device early warning method based on intelligent diagnosis according to claim 1, characterized in that, Multi-scale fault features are extracted using a multi-head temporal attention algorithm, specifically including: Set up multiple independent attention heads; Different time window scales are set for multiple independent attention heads; wherein, the time window scale is the length of the historical time step range that the attention mechanism focuses on; Generate a query vector based on the hidden state output of the Long Short-Term Memory network layer at the current moment; Based on the historical hidden state output sequence within the time scale window, generate the key vector matrix and the value vector matrix; Calculate the dot product of the query vector and the key vector matrix, divide by a preset scaling factor, perform masking, and apply the softmax function to obtain the weight distribution; The value vector matrix is ​​weighted and summed based on the weight distribution to determine the output vector of the independent attention head at the current time window scale. The output vectors of multiple independent attention heads are concatenated to output multi-scale fault features.

5. The device early warning method based on intelligent diagnosis according to claim 4, characterized in that, The multi-scale fault features are integrated into an adaptive feature fusion layer, specifically including: The current hidden state output of the Long Short-Term Memory Network is concatenated with the feature vector output by the multi-head temporal attention algorithm to obtain the concatenated feature vector. The concatenated feature vector is processed using the sigmoid activation function to generate a fused gated value vector. Based on the hidden state output of the long short-term memory network layer and the fusion gate value vector, the weighted hidden state features are determined. Based on the feature vector output by the multi-head temporal attention algorithm and the fusion gate value vector, the weighted attention features are determined. Based on the weighted hidden state features and the weighted attention features, an integrated feature is generated.

6. The device early warning method based on intelligent diagnosis according to claim 5, characterized in that, The fault prototype vector weights are updated based on changes in the operating conditions of the device under test, specifically including: based on determining a constraint inter-class distance to determine a prototype contrast loss; wherein, is a class prototype; is a prototype contrast loss, is a number of fault classes, is a first class sample feature mean, is a class sample feature output, is a margin of safety, is a class sample feature output, is a class sample feature output, is a Euclidean distance, is a number of samples of a class , is a set of i-th class samples, is a sample index, is a feature extraction network, is an n-th sample, denotes a max function; Introduce a pre-defined cross-entropy loss function; The total loss function is determined based on the prototype contrastive loss and the cross-entropy loss function: Determine the new batch data and historical batch data of the device to be tested; When the batch difference between the new batch feature distribution of the new batch data and the historical feature distribution of the historical batch data is greater than a preset difference threshold, the update rate is determined based on the value of the batch difference. The fault prototype vector weights are updated based on the update rate.

7. The device early warning method based on intelligent diagnosis according to claim 1, characterized in that, The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trend in the end-to-end fault diagnosis model, specifically including: The Euclidean distance between the current fused feature vector and the feature prototype vector of the end-to-end fault diagnosis model is measured and divided by the maximum norm of all category feature prototype vectors to determine the feature deviation. The fault probability distribution output by the end-to-end fault diagnosis model is calculated to determine the information entropy, thereby determining the classification uncertainty; wherein, the higher the entropy value of the classification uncertainty, the greater the uncertainty. The change in gradient entropy is determined by the absolute value of the difference between the weighted second-order gradient entropy eigenvalue at the current moment and the second-order gradient entropy eigenvalue at the previous moment. The feature deviation, classification uncertainty, and gradient entropy change are multiplied by preset weighting coefficients and then summed to determine the weighted composite value. The weighted composite value is input into the sigmoid function for nonlinear mapping to determine the confidence score.

8. A device early warning device based on intelligent diagnosis, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: A high-precision temperature sensor network is deployed to the key components of the device under test to obtain the raw temperature time-series data of the key components, and then sent to the preset central processing unit through a preset industrial IoT gateway. The central processing unit normalizes the original temperature time-series data to generate normalized temperature values. The weighted second-order gradient entropy characteristics are determined based on the normalized temperature values; An end-to-end fault diagnosis model is constructed. The input of the end-to-end fault diagnosis model is a normalized temperature value and a weighted second-order gradient entropy feature. The improved long short-term memory network layer of the end-to-end fault diagnosis model captures long-term dependencies. Multi-scale fault features are extracted by a multi-head temporal attention algorithm. The multi-scale fault features are output as integrated features by an adaptive feature fusion layer. The integrated features are output as a fault probability distribution by a fully connected classification layer. Update the fault prototype vector weights according to the changes in the operating conditions of the device under test; The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trends in the end-to-end fault diagnosis model. The real-time temperature data is sequentially generated into feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. An early warning is issued when the confidence score exceeds the confidence threshold and the failure probability in the failure probability distribution remains higher than the preset probability threshold. The construction of the improved long short-term memory network layer specifically includes: Calculate the average temperature gradient of the normalized temperature value within a sliding window that includes the current time point; The absolute value of the average temperature gradient is multiplied by a preset gating enhancement coefficient to generate a gradient gating term; The gradient gating term is added to the original input term of the forget gate to determine the forget gate input; The sigmoid activation function is applied to the enhanced forget gate input to determine the forget gate output; wherein the forget gate output determines the degree to which the cell state of the previous time step is preserved. 9.A non-volatile computer storage medium storing computer-executable instructions for implementing an intelligent diagnosis based device early warning, the computer-executable instructions comprising: The computer-executable instructions are set as follows: A high-precision temperature sensor network is deployed to the key components of the device under test to obtain the raw temperature time-series data of the key components, and then sent to the preset central processing unit through a preset industrial IoT gateway. The central processing unit normalizes the original temperature time-series data to generate normalized temperature values. The weighted second-order gradient entropy characteristics are determined based on the normalized temperature values; An end-to-end fault diagnosis model is constructed. The input of the end-to-end fault diagnosis model is a normalized temperature value and a weighted second-order gradient entropy feature. The improved long short-term memory network layer of the end-to-end fault diagnosis model captures long-term dependencies. Multi-scale fault features are extracted by a multi-head temporal attention algorithm. The multi-scale fault features are output as integrated features by an adaptive feature fusion layer. The integrated features are output as a fault probability distribution by a fully connected classification layer. Update the fault prototype vector weights according to the changes in the operating conditions of the device under test; The confidence score is determined based on the feature deviation, classification uncertainty, and gradient entropy change trends in the end-to-end fault diagnosis model. The real-time temperature data is sequentially generated into feature vectors, which are then input into the trained end-to-end fault diagnosis model to obtain the fault probability distribution. An early warning is issued when the confidence score exceeds the confidence threshold and the failure probability in the failure probability distribution remains higher than the preset probability threshold. The construction of the improved long short-term memory network layer specifically includes: Calculate the average temperature gradient of the normalized temperature value within a sliding window that includes the current time point; The absolute value of the average temperature gradient is multiplied by a preset gating enhancement coefficient to generate a gradient gating term; The gradient gating term is added to the original input term of the forget gate to determine the forget gate input; The sigmoid activation function is applied to the enhanced forget gate input to determine the forget gate output; wherein the forget gate output determines the degree to which the cell state of the previous time step is preserved.

Images