A front-end device multi-dimensional fault determination method and system

Abstract

The application provides a front-end device multi-dimensional fault judgment method and system, and relates to the technical field of fault diagnosis. The method comprises the following steps: collecting multi-dimensional original monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity working conditions in a current operation cycle; performing time sequence benchmark alignment and transient interference adaptive filtering processing on the multi-dimensional original monitoring data streams, extracting steady-state electrical characteristic baselines and nonlinear dynamic evolution trajectories of each monitoring channel, and obtaining a multi-dimensional time sequence feature fusion set; performing feature decoupling and cross-domain mapping processing on the multi-dimensional time sequence feature fusion set, and obtaining hardware electrical drift components, communication link signal-to-noise ratio attenuation components and micro-environment disturbance components; and taking the hardware electrical drift components, the communication link signal-to-noise ratio attenuation components and the micro-environment disturbance components as inputs, and constructing a dynamic directional surrounding geometric body representing the evolution trajectories of the components. The application improves the operation reliability and production continuity of the equipment.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method and system for multi-dimensional fault determination of front-end equipment. Background Technology

[0002] With the deepening of industrial digital transformation, industrial edge front-end equipment, as the core node for data acquisition, real-time processing and command issuance on the production site, is directly related to the stable operation of the entire intelligent manufacturing system. In typical harsh industrial scenarios such as hot rolling of steel and non-ferrous metal smelting, edge computing gateways, data acquisition terminals and other equipment deployed on the production site are subjected to transient electromagnetic pulses generated by the start and stop of high-power motors, thermal and humid shocks from the alternating effects of high-temperature steel billet radiation and cooling water spray, and mechanical vibrations generated by equipment operation. These devices are prone to hidden faults that are difficult to detect, such as hardware parameter drift and communication link attenuation.

[0003] These types of latent faults initially manifest as intermittent anomalies that do not trigger hard shutdowns of equipment. However, they can lead to decreased data acquisition accuracy and delayed control commands, gradually accumulating until they eventually cause production accidents. For example, the edge computing gateway on a hot rolling production line experienced increased voltage ripple due to the slow aging of capacitors in the main control power supply circuit. This, coupled with strong electromagnetic interference from the on-site rolling mill frequency converter, caused intermittent communication interruptions and data packet loss. Traditional fault diagnosis methods rely solely on a single communication bit error rate threshold for alarms, making it difficult to distinguish the coupled effects of three types of fault components: hardware electrical drift, communication link degradation, and micro-environmental disturbances. This makes it difficult to accurately pinpoint the root cause of the fault and to predict the trend of equipment deterioration. Consequently, maintenance personnel must systematically check power supply modules, communication interfaces, and environmental sensing units, resulting in an average fault repair time exceeding 4 hours and impacting production continuity. Summary of the Invention

[0004] This invention provides a method and system for multi-dimensional fault determination of front-end equipment, which improves the reliability of equipment operation and production continuity.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for multi-dimensional fault determination of a front-end device, the method comprising:

[0007] Collect multi-dimensional raw monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity conditions during the current operating cycle; perform time-series benchmark alignment and transient interference adaptive filtering on the multi-dimensional raw monitoring data streams, extract the steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel, and obtain a multi-dimensional time-series feature fusion set;

[0008] Feature decoupling and cross-domain mapping are performed on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components. Using the hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamic oriented enclosing geometry representing the evolution trajectory of each component is constructed.

[0009] The spatiotemporal collaborative association weights are determined by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary is calculated to obtain the multidimensional feature deviation.

[0010] The spatiotemporal collaborative correlation weights and multidimensional feature deviations are superimposed by tensor mapping to construct a multidimensional fault feature correlation matrix; the multidimensional fault feature correlation matrix is ​​input into a preset fault mode identification strategy, and a topological state traversal operation is performed to obtain a feature-coupled topological state sequence.

[0011] By performing node similarity matching and causal chain deduction between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph, the fault location and classification results are obtained.

[0012] Secondly, a multi-dimensional fault determination system for front-end devices includes:

[0013] The acquisition module is used to acquire multi-dimensional raw monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity conditions during the current operating cycle; it performs time-series benchmark alignment and transient interference adaptive filtering on the multi-dimensional raw monitoring data streams, extracts the steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel, and obtains a multi-dimensional time-series feature fusion set;

[0014] The module is used to perform feature decoupling and cross-domain mapping processing on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components; using the hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamic oriented enclosing geometry representing the evolution trajectory of each component is constructed;

[0015] The calculation module is used to determine the spatiotemporal collaborative association weight by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and to calculate the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary to obtain the multidimensional feature deviation.

[0016] The mapping module is used to superimpose the spatiotemporal collaborative correlation weights and multidimensional feature deviations through tensor mapping to construct a multidimensional fault feature correlation matrix; the multidimensional fault feature correlation matrix is ​​input into a preset fault mode identification strategy, and a topological state traversal operation is performed to obtain a feature-coupled topological state sequence.

[0017] The judgment module is used to perform node similarity matching and causal chain inference between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph to obtain fault location and classification judgment results.

[0018] Thirdly, a computing device includes:

[0019] One or more processors;

[0020] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0021] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] By employing time-series benchmark alignment and adaptive filtering of transient interference from multi-dimensional raw monitoring data streams, combined with feature decoupling and dynamic orientation enclosing geometry modeling, as well as tensor mapping superposition of spatiotemporal collaborative correlation weights and multi-dimensional feature deviations, topological state traversal operations, and node similarity matching and causal chain inference techniques from industrial equipment fault knowledge graphs, this approach overcomes the technical problems of high data noise and severe time-series offset under strong electromagnetic interference and alternating temperature and humidity conditions, severe feature coupling and difficulty in distinguishing the evolution trends of different fault components in traditional fault determination methods, and inaccurate fault location, unclear classification, and difficulty in predicting equipment condition deterioration trends caused by single threshold alarms. This results in an accurate multi-dimensional time-series feature fusion set, enabling independent representation and dynamic tracking of fault components, accurate root cause location and classification of faults in industrial edge front-end equipment, shortening fault repair time, and improving equipment operational reliability and production continuity. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a multi-dimensional fault determination method for front-end devices provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a multi-dimensional fault determination system for front-end devices provided by an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] like Figure 1 As shown, an embodiment of the present invention proposes a multi-dimensional fault determination method for front-end devices, the method comprising the following steps:

[0028] Step 1: Collect multi-dimensional raw monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity conditions during the current operating cycle; perform time-series benchmark alignment and transient interference adaptive filtering on the multi-dimensional raw monitoring data streams, extract the steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel, and obtain a multi-dimensional time-series feature fusion set;

[0029] Step 2: Perform feature decoupling and cross-domain mapping processing on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift component, communication link signal-to-noise ratio attenuation component and micro-environmental disturbance component; use the hardware electrical drift component, communication link signal-to-noise ratio attenuation component and micro-environmental disturbance component as input to construct a dynamic oriented enclosing geometry that characterizes the evolution trajectory of each component;

[0030] Step 3: Determine the spatiotemporal collaborative association weight by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and calculate the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary to obtain the multidimensional feature deviation.

[0031] Step 4: Tensor mapping is performed to superimpose the spatiotemporal collaborative association weights and the multidimensional feature deviation to construct a multidimensional fault feature association matrix; the multidimensional fault feature association matrix is ​​input into the preset fault mode identification strategy, and the topological state traversal operation is performed to obtain the feature-coupled topological state sequence.

[0032] Step 5: Perform node similarity matching and causal chain deduction between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph to obtain fault location and classification results.

[0033] In this embodiment of the invention, by employing multi-dimensional raw monitoring data stream acquisition and time-series benchmark alignment, transient interference adaptive filtering processing, combined with feature decoupling and cross-domain mapping, dynamic orientation enclosing geometry construction, spatiotemporal collaborative association weight and multi-dimensional feature deviation calculation, tensor mapping superposition to construct fault feature association matrix, topological state traversal operation, and industrial equipment fault knowledge graph node similarity matching and causal chain inference, this invention overcomes the technical problems of high data noise and severe time-series offset under strong electromagnetic interference and alternating temperature and humidity conditions, severe multi-source feature coupling and inability to distinguish the evolution trend of different fault components in traditional fault determination methods, and the fuzzy fault root cause location, unclear hierarchical determination, and inability to track the equipment condition deterioration process caused by single threshold alarms. This results in an accurate multi-dimensional time-series feature fusion set, enabling independent representation and dynamic tracking of three types of fault components: hardware electrical, communication link, and microenvironment. This completes the root cause location and hierarchical determination of faults in industrial edge front-end equipment, shortens fault investigation time, and improves equipment operation stability and production continuity.

[0034] In a preferred embodiment of the present invention, step 1 above may include:

[0035] Step 1.1: Real-time reading of voltage ripple sampling values ​​and current load fluctuation sampling values ​​of the main control power supply circuit to obtain the raw sequence of hardware electrical telemetry; synchronously capturing the bit error rate count, carrier signal-to-noise ratio fluctuation value, and link handshake status identifier of the physical layer of the industrial communication interface to obtain the raw sequence of communication link status. Specifically, this includes: obtaining the voltage ripple sampling value of the main control power supply circuit through a high-precision voltage sampling resistor connected in parallel across the positive and negative terminals of the power supply circuit in conjunction with a 16-bit analog-to-digital converter. The high-precision voltage sampling resistor is a low-temperature drift metal film resistor with a resistance of 10kΩ and a temperature coefficient of ±5ppm / ℃. The 16-bit analog-to-digital converter adopts a successive approximation architecture, with an input voltage range of 0V to 5V and a conversion time of 1μs. The formula for converting the raw digital quantity output by the analog-to-digital converter into the actual ripple voltage is as follows:

[0036] ;

[0037] In the formula, This represents the actual voltage ripple value, in millivolts. This is the 16-bit raw digital value output by the analog-to-digital converter. This is the reference voltage for the analog-to-digital converter, with a value of 5V. The voltage gain coefficient of the front-end sampling circuit is determined by the voltage division ratio of the sampling resistor and the amplification factor of the operational amplifier. In this embodiment, it is set to 0.02. The current load fluctuation sampling value is obtained through a closed-loop Hall current sensor connected in series with the positive terminal of the power supply circuit in conjunction with a 16-bit analog-to-digital converter. The measurement range of the closed-loop Hall current sensor is 0A to 20A, with a linearity of 0.1% and a response time of 1μs. The formula for converting the raw digital quantity output by the analog-to-digital converter into the actual load current is as follows:

[0038] ;

[0039] In the formula, This is the actual load current value, in amperes. The Hall sensor's full-scale measurement range is set to 20A, and the sampling frequency is set to 1kHz, meaning an analog-to-digital conversion operation is performed every 1ms. The analog-to-digital converter converts the analog signal into a 16-bit binary digital signal, which is then transmitted to the general-purpose input / output port of the device's main control chip via a 16-bit parallel data bus. The main control chip performs the aforementioned physical quantity conversion operations on the received digital signal and arranges them according to the chronological order of acquisition time to obtain the original hardware electrical telemetry sequence. The bit error rate counter, carrier signal-to-noise ratio fluctuation value, and link handshake status identifier of the industrial communication interface physical layer are read through the interface chip's built-in status register. The industrial communication interface chip has a built-in 32-bit bit error rate counter, a 16-bit carrier signal-to-noise ratio register, and an 8-bit link status register. The reading frequency is set to 100Hz, meaning a register read operation is performed every 10ms. The formula for converting the original register values ​​into the actual bit error rate is:

[0040] ;

[0041] In the formula, This represents the actual bit error rate, which is dimensionless. This is the original read value of the bit error rate counter; The length of a single frame of data is 1518 bytes, or 12144 bits, in this embodiment. This is the register read frequency, with a value of 100Hz. The physical layer data transmission rate of the interface is set to 100Mbps. The main control chip establishes a communication connection with the industrial communication interface chip through the serial peripheral interface. The clock frequency of the serial peripheral interface is set to 10MHz, the data width is 8 bits, and master-slave communication is adopted. The main control chip sends a register read instruction to the interface chip. The instruction includes the slave device address, the register start address, and the read length. After receiving the instruction, the interface chip returns the data in the corresponding register to the main control chip. The main control chip performs physical quantity conversion operations on the read values ​​and arranges them in the order of the read time to obtain the original sequence of communication link status.

[0042] Step 1.2: Collect the output sampling values ​​of the temperature sensing unit, humidity sensing unit, and triaxial vibration accelerometer on the inner wall of the protective chassis to obtain the original sequence of microenvironment sensing; simultaneously extract the interrupt response delay record, memory stack allocation fragmentation rate, and task scheduling blocking duration of the embedded real-time kernel to obtain the original sequence of kernel running logs. Specifically, the temperature sensing unit and humidity sensing unit on the inner wall of the protective chassis are integrated into the same sensing chip. The temperature measurement range of the sensing chip is -40℃ to 125℃, and the measurement resolution is 0.01℃; the humidity measurement range is 0%RH to 100%RH, and the measurement resolution is 0.01%RH. The sensor chip is connected to the device's main control chip via an integrated circuit bus. The clock frequency of the integrated circuit bus is set to 100kHz, using 7-bit slave addressing, and the sampling frequency is set to 10Hz, meaning that temperature and humidity data are read every 100ms. The triaxial vibration accelerometer has a measurement range of ±16g, a resolution of 16 bits, and a bandwidth of 1kHz. The triaxial vibration accelerometer is connected to the device's main control chip via a serial peripheral interface. The clock frequency of the serial peripheral interface is set to 1MHz, the data bit width is 8 bits, and the sampling frequency is set to 500Hz, meaning that acceleration data is read every 2ms. The main control chip periodically sends read commands to the sensor chip. After receiving the command, the sensor chip initiates internal analog-to-digital conversion, converting the analog signals of temperature, humidity, and triaxial vibration acceleration into digital signals, and returns the conversion result to the main control chip. The received digital signals are temporarily stored and arranged in chronological order of acquisition time to obtain the original sequence of microenvironment sensing. The interrupt response latency record, memory stack allocation fragmentation rate, and task scheduling blocking time of the embedded real-time kernel are extracted through the system call interface provided by the kernel. The interrupt response latency record is the time interval from the generation of the interrupt request signal to the start of the interrupt service routine, in microseconds. The memory stack allocation fragmentation rate is the ratio of the number of unused memory blocks in the kernel heap memory to the total number of memory blocks. The task scheduling blocking time is the waiting time for a task to transition from the ready state to the running state, in microseconds. The extraction frequency is set to 10Hz, that is, a call operation is performed every 100ms. The main control chip calls the kernel system interface to obtain the current value of the corresponding parameter, temporarily stores the obtained value, and arranges it in chronological order of extraction time to obtain the original sequence of kernel running log.

[0043] Step 1.3 involves injecting absolute timestamps generated by a unified hardware clock source into the four types of raw sequences, performing cross-domain data frame encapsulation and buffer alignment to obtain a multi-dimensional raw monitoring data stream. Specifically, the unified hardware clock source uses a built-in temperature-controlled crystal oscillator with an output frequency of 1MHz, a frequency stability of ±0.1ppm, and a temperature stability range of -40℃ to 85℃. The clock counter is a 32-bit incrementing counter that continuously counts the output pulses of the temperature-controlled crystal oscillator, with a counting range of 0 to 4294967295. Simultaneously with the acquisition or extraction of each raw data point, the main control chip reads the current clock counter value and calculates the absolute timestamp corresponding to that data point. The formula for calculating the absolute timestamp is:

[0044] ;

[0045] in, This is an absolute timestamp, in seconds. This is the current count value of the clock counter; To unify the output frequency of the hardware clock source, measured in Hertz, the main control chip appends the calculated absolute timestamp to the header of the corresponding original data point, completing the timestamp injection operation. Cross-domain data frame encapsulation encapsulates each data point with an absolute timestamp from the four types of original sequences according to a unified frame format. The total frame format length is 64 bytes, containing a 1-byte frame header, a 1-byte data type identifier, an 8-byte absolute timestamp field, a 1-byte data length field, a 51-byte data content field, and a 2-byte cyclic redundancy check field. The frame header is fixed at 0xAA, used to identify the start position of the data frame. The data type identifier distinguishes between hardware electrical telemetry data, communication link status data, microenvironment sensor data, and kernel runtime log data. The absolute timestamp field stores a 64-bit double-precision floating-point absolute timestamp, and the data length field represents the data content field... The data content field stores the values ​​of the original data points, and the cyclic redundancy check field is calculated using the CRC-16 algorithm. The calculation scope includes the data type identifier, absolute timestamp field, data length field, and data content field. Cache alignment involves storing the encapsulated data frames in a circular buffer according to the order of their absolute timestamps. The size of the circular buffer is 1024 data frames, and a first-in-first-out read / write mechanism is adopted. When data frames with different sampling frequencies arrive in the buffer, the main control chip compares the absolute timestamp of the newly arrived data frame with the absolute timestamp of the data frames already in the buffer, and inserts the new data frame into the appropriate storage location. This ensures that data frames of different dimensions collected at the same time are in adjacent storage locations in the buffer. When the buffer is full, the earliest written data frame is overwritten, ultimately resulting in a multi-dimensional original monitoring data stream with consistent timing and uniform format.

[0046] In this embodiment of the invention, by adopting a synchronous acquisition mechanism for multi-dimensional raw data, combined with the injection of absolute timestamps generated by a unified hardware clock source, unified encapsulation of cross-domain data frames, and timestamp-based cache alignment technology, the technical problems of asynchronous timing of multi-source data under strong electromagnetic interference and alternating temperature and humidity conditions, data misalignment caused by transmission link jitter, and inability to correlate and analyze data of different dimensions are overcome, thereby obtaining a multi-dimensional raw monitoring data stream with consistent timing and unified format.

[0047] In a preferred embodiment of the present invention, step 1 above may include:

[0048] Step 1.4: Based on the absolute timestamps in the multi-dimensional raw monitoring data stream, resampling and phase compensation are performed on the discrete sampling sequences of each channel to eliminate the timing offset caused by transmission link jitter, resulting in aligned data sequences for each channel. Specifically, this includes parsing discrete sampling sequences from four channels—hardware electrical telemetry, communication link status, micro-environment sensing, and kernel operation logs—from the multi-dimensional raw monitoring data stream. Each sequence contains several data points with absolute timestamps. A unified resampling frequency of 100Hz is determined, and a resampling time point sequence with equal intervals within the current operating cycle is generated. The formula for calculating the resampling time point is:

[0049] ;

[0050] In the formula, For the first Each resampling time point is represented in seconds. This is the start time of the current running cycle, in seconds. This is the index of the resampling point, with a value ranging from 0 to... -1, This represents the total number of resampling points within the current running cycle. The resampling time interval, in seconds, is set to 0.01. Cubic spline interpolation is performed on the original discrete sampling sequence of each channel to construct a piecewise cubic polynomial function. Each interval between adjacent original sampling points corresponds to a cubic polynomial. The first and second derivatives of the polynomial are continuous at the endpoints of the interval. Natural boundary conditions are used, with the second derivative at the beginning and end of the sequence being zero. The coefficients of the cubic polynomial for each interval are obtained by solving a tridiagonal linear system of equations. These coefficients are then substituted into the resampling time points to calculate the corresponding real-valued interpolation sequence. The real number sequence obtained by cubic spline interpolation Perform Hilbert transform to construct analytic signal ,in This represents the Hilbert transform, and the operation is implemented through temporal convolution with a convolution kernel of... The real part of the analytic signal is taken as the value of the interpolated real part. The imaginary part is used as the interpolation imaginary part value. Calculate the time difference between each original sampling point and the most recent resampling time point, and calculate the phase offset based on the time difference. The formula for calculating the phase offset is:

[0051] ;

[0052] In the formula, For the first Phase offset corresponding to each original sampling point; To determine the center frequency of the corresponding channel signals, the hardware electrical telemetry channel is set to 50Hz, the communication link status channel is set to 10Hz, the microenvironment sensing channel is set to 1Hz, and the kernel running log channel is set to 1Hz. For the first The time difference between each original sampling point and the most recent resampling time point is used to correct the interpolated data value based on the phase offset. The correction formula is as follows:

[0053] ;

[0054] In the formula, For the revised first Values ​​of each resampled point; The interpolated first The real part value of each resampled point; The interpolated first The imaginary part values ​​of each resampling point are used to arrange the corrected interpolated data values ​​of each channel according to the order of the resampling time points, resulting in the aligned data sequence of each channel.

[0055] Step 1.5: Input the aligned data sequences of each channel into the adaptive threshold sliding filter window. Dynamically adjust the filter window width and attenuation coefficient based on the transient peak waveform characteristics of the strong electromagnetic pulse and the thermal inertia characteristics of ambient temperature changes to filter out non-faulty high-frequency noise and step interference, obtaining a purified data sequence. Specifically, this includes: initializing the adaptive threshold sliding filter window, setting the initial window width to 7 sampling points and the initial attenuation coefficient to 0.5. Input the aligned data sequences of each channel into the filter window sequentially, and calculate the instantaneous rate of change of the current sampling point within the window. The formula for calculating the instantaneous rate of change is:

[0056] ;

[0057] In the formula, For the first The instantaneous rate of change of each sampling point; For the first The values ​​of each sampling point; For the first -1 sample point values; Let be the sampling time interval, with a value of 0.01 seconds. Calculate the standard deviation of all data points within the current sliding window. The formula for calculating the standard deviation is:

[0058] ;

[0059] In the formula, The standard deviation of the data within the window; This is the current filter window width; This represents the value of the p-th sampling point within the window; The filter window is set to the average of all sampling points within the window. A preset peak threshold is 5 times the standard deviation of the data within the window, and a preset gradient threshold is 0.2 times the standard deviation of the data within the window. When the instantaneous rate of change is greater than the peak threshold, the current sampling point is determined to be a transient peak interference caused by a strong electromagnetic pulse. The filter window width is adjusted to 3 sampling points, and the attenuation coefficient is adjusted to 0.1. When the instantaneous rate of change is less than the gradient threshold, the current data change is determined to be a slow gradient caused by environmental temperature change. The filter window width is adjusted to 15 sampling points, and the attenuation coefficient is adjusted to 0.8. When the instantaneous rate of change is between the peak threshold and the gradient threshold, the current filter window width and attenuation coefficient remain unchanged. Based on the adjusted filter window width and attenuation coefficient, a weighted moving average filter operation is performed on the data points within the window. The formula for calculating the weighted moving average is:

[0060] ;

[0061] In the formula, For the filtered first The values ​​of each sampling point; This represents the weight of the p-th sampling point within the window. Sampling points closer to the current sampling point have a larger weight. The weight calculation formula is as follows: , Given the current attenuation coefficient, the sliding window moves forward by one sampling point each time, repeating the above calculation process, traversing all sampling points of all channels, completing the filtering process, and obtaining the purified data sequence.

[0062] Step 1.6: Perform low-pass smoothing on the purified data sequence and extract the steady-state electrical characteristic baseline of each monitoring channel within the preset normal operating range; calculate the residual sequence between the purified data sequence and the steady-state electrical characteristic baseline, specifically including: performing Butterworth low-pass smoothing on the purified data sequence, setting the Butterworth filter order to 4th order and the cutoff frequency to 5Hz, and calculating the simulated Butterworth low-pass filter, whose squared amplitude response function is:

[0063] ;

[0064] in, This is the frequency response of the analog filter; To simulate angular frequency; The cutoff angular frequency; Let be the filter order, with a value of 4. The analog filter is converted to a digital filter using the bilinear transform method. The formula for the bilinear transform is:

[0065] ;

[0066] In the formula, It is a complex frequency variable; The sampling period is 0.01 seconds. Let z be the z-transform variable. After obtaining the system function of the digital filter, convert it into a difference equation form and perform point-by-point filtering calculations on the purified data sequence. Extract the smoothed data within a preset normal operating range, which is the data range of 10 consecutive operating cycles after the equipment is powered on and running stably. Calculate the moving average of the data within this range, with the sliding window size set to 20 sampling points. The formula for calculating the moving average is:

[0067] ;

[0068] In the formula, Let be the baseline value of the steady-state electrical characteristics at the i-th point; For the smoothed first The values ​​of each sampling point are used to calculate the difference between each sampling point in the purified data sequence and the corresponding steady-state electrical characteristic baseline point, thus obtaining the residual sequence. The formula for calculating the residual sequence is as follows:

[0069] ;

[0070] In the formula, The th in the residual sequence The value of each point; To clean the data sequence of the first The value of each point; The first in the steady-state electrical characteristic baseline The value of each point.

[0071] Step 1.7: Perform gradient differencing and nonlinear envelope tracking on the residual sequence to extract the nonlinear dynamic evolution trajectory characterizing the equipment condition degradation trend; concatenate and normalize the feature vectors of the steady-state electrical characteristic baseline and the nonlinear dynamic evolution trajectory according to the channel dimension to obtain a multi-dimensional time series feature fusion set. Specifically, this includes: performing first-order gradient differencing on the residual sequence to obtain the rate of change sequence of the residual sequence. The calculation formula for the first-order gradient differencing is:

[0072] ;

[0073] In the formula, For the first The rate of change values ​​at the nth point are used to perform trend extraction processing on the rate of change sequence using sliding window extremum detection and cubic spline interpolation. Local maxima and minima are extracted using the sliding window comparison method, with the sliding window size set to 5 sampling points. For the nth point in the rate of change sequence... A point is identified as a local maximum if its value is greater than the values ​​of all other points within the window; otherwise, it is identified as a local minimum. Adjacent local maxima are connected by cubic spline curves to obtain the upper envelope of the residual sequence; adjacent local minima are connected by cubic spline curves to obtain the lower envelope of the residual sequence. The average values ​​of the upper and lower envelopes at each sampling point are calculated to obtain the nonlinear dynamic evolution trajectory characterizing the equipment's deterioration trend.

[0074] The envelope average reflects the fluctuation center shift trend of the residual signal. Verified using historical fault samples, its output has a Pearson correlation coefficient of 0.92 with the capacitance decay rate and a Pearson correlation coefficient of 0.89 with the communication bit error rate rise rate, effectively characterizing the slow degradation process of equipment status. The steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel are concatenated as feature vectors along the channel dimension. Each channel corresponds to two feature dimensions. The feature dimensions of all channels are concatenated in the order of hardware electrical telemetry, communication link status, microenvironment sensing, and kernel operation log to obtain the initial feature vector. Robust dimensional normalization mapping based on the rated baseline is performed on the initial feature vector. The rated health value of each feature dimension is used as the normalization neutral point, and scaling is performed using the interquartile range to eliminate the influence of outliers and dimensional differences. The normalization calculation formula is:

[0075] ;

[0076] In the formula, For the normalized first One eigenvalue; The first eigenvector in the initial eigenvector One eigenvalue; For the first The rated health values ​​corresponding to each feature dimension are taken from the normal operating reference values ​​measured in the equipment's factory calibration test, such as the rated voltage ripple value of 50mV and the rated communication bit error rate value of 1e-6. For the first The interquartile range of all samples in each feature dimension is equal to the difference between the upper and lower quartiles. The normalized feature vectors are arranged in chronological order to obtain a multi-dimensional temporal feature fusion set.

[0077] In this embodiment of the invention, by employing unified resampling and phase offset correction processing based on cubic spline interpolation, combined with adaptive threshold sliding filtering based on instantaneous rate of change, Butterworth low-pass smoothing baseline extraction, residual gradient difference and nonlinear envelope tracking, and robust feature normalization mapping based on median and interquartile range, the technical problems of time series offset of multi-source data under strong electromagnetic interference and alternating temperature and humidity conditions, severe non-fault noise interference, difficulty in extracting equipment condition degradation trends, feature distortion caused by outliers, and difficulty in fusion analysis due to inconsistent feature dimensions of different dimensions are overcome. Thus, a multi-dimensional time series feature fusion set with consistent time series, low noise interference, including steady-state operation characteristics and dynamic degradation trends of equipment, and strong resistance to outliers is obtained.

[0078] In a preferred embodiment of the present invention, step 2 above may include:

[0079] Step 2.1: Map the multi-dimensional temporal feature fusion set to a three-dimensional orthogonal feature space. This is done by constructing a multi-dimensional feature covariance matrix and solving for the principal component eigenvectors to determine the principal axis distribution direction of the feature space. Specifically, the multi-dimensional temporal feature fusion set is a two-dimensional matrix structure. Rows correspond to time steps, and columns correspond to feature dimensions. Each time step includes steady-state and dynamic evolution features from four channels: hardware electrical telemetry, communication link status, micro-environment sensing, and kernel operation logs, totaling eight feature dimensions. The current running cycle duration is 30 seconds, the resampling frequency is 100 Hz, and the multi-dimensional temporal feature fusion set is a 3000-row × 8-column two-dimensional matrix. Mapping the multi-dimensional temporal feature fusion set to a three-dimensional orthogonal feature space and constructing the multi-dimensional feature covariance matrix is ​​described by the following formula:

[0080] ;

[0081] In the formula, It is an 8×8 multidimensional feature covariance matrix, where the diagonal elements are the variances of each feature dimension, and the off-diagonal elements are the covariances between different feature dimensions; The total number of time steps in the multi-dimensional temporal feature fusion set, with a value of 3000; For the first An 8-dimensional feature vector at each time step; This is an 8-dimensional mean vector of the feature vectors at all time steps, where each element is the average value of the corresponding feature dimension across all time steps. The covariance matrix is ​​solved using the Jacobi iteration method. The eigenvalues ​​and eigenvectors are obtained, and the iteration terminates when the absolute value of all off-diagonal elements is less than 1e-6. The solved eigenvalues ​​are arranged in descending order to obtain the eigenvalue sequence. For each eigenvalue, the corresponding eigenvector is normalized by dividing the eigenvector by its magnitude to obtain a unit eigenvector. The three largest eigenvalues ​​are then selected. , , corresponding unit eigenvector , , As the principal axis direction vectors of the three-dimensional orthogonal feature space, the three principal axis direction vectors are pairwise orthogonal and have a magnitude of 1, forming the standard orthogonal basis of the feature space.

[0082] Step 2.2 involves performing orthogonal projection of the feature space and channel crosstalk stripping based on the main axis distribution direction. This decouples and outputs the hardware electrical drift component, which independently represents the slow shift of the main control power supply circuit parameters; the communication link signal-to-noise ratio attenuation component, which independently represents the degradation of the underlying signal quality of the communication interface; and the micro-environmental disturbance component, which independently represents the superposition of thermal and humidity changes and mechanical vibrations inside and outside the chassis. Specifically, this includes performing orthogonal projection of the feature space and channel crosstalk stripping based on the main axis distribution direction. This involves fusing the multi-dimensional temporal features into each 8-dimensional feature vector. Performing dot products with the three 8-dimensional unit principal axis direction vectors respectively yields three independent scalar projection components. The formula for calculating orthogonal projection is:

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, For the first The projection component of each time step in the first principal axis direction; For the first The projection component of each time step in the second principal axis direction; For the first The projection components at each time step along the third principal axis are calculated, and the covariance matrix between the three projection components is determined. Due to the orthogonality of principal component analysis, this covariance matrix is ​​a diagonal matrix, with all off-diagonal elements being zero, indicating that there is no linear correlation between the three components, thus achieving complete isolation of channel crosstalk. The physical meaning of the components is verified by labeling historical fault data. The first principal axis corresponds to the dimension with the largest data variance, and its correlation with the changing trends of voltage ripple and current load in the main control power supply circuit is greater than 0.9, with the decoupled output being the hardware electrical drift component. The second principal axis corresponds to the dimension with the second largest data variance, and its correlation with the changing trends of communication bit error rate and carrier signal-to-noise ratio is greater than 0.85, with the decoupled output being the communication link signal-to-noise ratio attenuation component. The third principal axis corresponds to the dimension with the third largest data variance, and its correlation with the changing trends of temperature, humidity, and vibration acceleration inside the chassis is greater than 0.8, with the decoupled output being the micro-environmental disturbance component.

[0087] Step 2.3: Calculate the spatial discrete distribution boundaries of the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component within the sliding time window. Using the time-series extreme point set of the three components as spatial vertex constraints, perform coordinate axis rotation and alignment along the principal component eigenvector direction to obtain the initial oriented bounding body encompassing the discrete distribution of features within the current running cycle. Specifically, this includes: setting the length of the sliding time window to 1 / 3 of the current running cycle duration, i.e., 10 seconds, corresponding to 1000 consecutive time steps; the sliding step size is 1 time step, consistent with the resampling frequency; within the sliding time window, extract the maximum and minimum points of the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component, respectively. Each point contains the component value and the corresponding timestamp. There are a total of 6 extreme points for the three components, forming a time-series extreme point set. Using the time-series extreme point set of the three components as spatial vertex constraints, perform coordinate axis rotation and alignment along the principal component eigenvector direction to construct the initial oriented bounding body. The centroid coordinates of the initial oriented bounding body are the average value of the three components within the sliding time window, calculated using the following formula:

[0088] ;

[0089] In the formula, The coordinates of the three-dimensional centroid of the initial oriented bounding volume; This represents the average value of the hardware electrical drift components within the sliding time window; This represents the average value of the signal-to-noise ratio attenuation component of the communication link within the sliding time window; This represents the average value of the microenvironmental disturbance components within the sliding time window. The three principal axis lengths of the initial orientation bounding volume are the differences between the maximum and minimum values ​​of the corresponding components, calculated using the following formula:

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, This is the length of the first spindle; This is the length of the second main axis; This is the length of the third spindle; , These represent the maximum and minimum values ​​of the hardware electrical drift component within the sliding time window, respectively. , These represent the maximum and minimum values ​​of the signal-to-noise ratio attenuation component of the communication link within the sliding time window; , These represent the maximum and minimum values ​​of the microenvironmental disturbance components within the sliding time window. The spatial rotation attitude of the initial orientation bounding volume is determined by a 3×3 orthogonal rotation matrix composed of the eigenvectors of the three principal components. The column vectors of the rotation matrix are respectively... , , Ensure that the three principal axes of the bounding volume are perfectly aligned with the principal axes of the feature space. Calculate the coordinates of the eight vertices of the initial oriented bounding volume; the coordinates of the eight vertices in the local coordinate system are... The transformation to the global feature space coordinate system is achieved through a rotation matrix and the centroid coordinates. The transformation formula is as follows: ,in These are the vertex coordinates in the local coordinate system. Let be a rotation matrix. Given the vertex coordinates in the global coordinate system, verify that all feature points within the sliding time window are located inside the bounding volume. That is, after each feature point is transformed to the local coordinate system, the absolute value of each coordinate component is less than or equal to half the length of the corresponding principal axis, thus obtaining the initial oriented bounding volume that encloses the discrete distribution of features within the current running cycle.

[0094] Step 2.4 involves continuously updating the centroid coordinates, principal axis length ratios, and spatial rotation attitude of the initial oriented bounding volume based on the temporal evolution of the three components, resulting in a dynamically oriented bounding geometry. Specifically, this includes: moving the sliding time window forward at a resampling frequency, updating the three component data within the sliding time window at each time step; for each new time step, removing the oldest data point within the sliding time window and adding the latest data point; and calculating the average value of the three components within the sliding time window using an incremental method. The update formula is as follows: ,in This is the updated average. This is the average value before the update. For newly added data points, For the data points to be removed, The sliding time window length is set to 1000. An incremental method is used to update the maximum and minimum values ​​of the three components. When the removed data point is the current maximum value, the remaining data points in the window are re-traversed to find the new maximum value. When the removed data point is the current minimum value, the remaining data points in the window are re-traversed to find the new minimum value. Otherwise, the newly added data point is compared with the current maximum and minimum values, and the maximum or minimum value is updated. An incremental principal component analysis algorithm is used to update the principal component eigenvectors. The eigenvalues ​​and eigenvectors of the covariance matrix are updated using rank-1 correction, avoiding recalculation of the entire covariance matrix each time. Based on the updated parameters, the centroid coordinates, principal axis length ratio, and spatial rotation attitude of the initial oriented bounding volume are continuously adjusted. The update frequency of the centroid coordinates is consistent with the resampling frequency, which is 100Hz. The principal axis length ratio is adjusted in real time according to the extreme value changes of the three components, reflecting the different evolution rates of different fault components. The spatial rotation attitude is fine-tuned according to the changes in the principal component eigenvectors. When the angle change of the eigenvector is less than 0.1 degrees, the rotation attitude is not adjusted; when the angle change is greater than or equal to 0.1 degrees, the rotation matrix is ​​updated and the vertex coordinates of the bounding volume are recalculated. The change in the centroid coordinates of the bounding volume between adjacent time steps does not exceed 1% of the current principal axis length, the change in the principal axis length does not exceed 5% of the current principal axis length, and the change in the rotation angle does not exceed 0.1 degrees, ensuring a continuous and smooth evolution process for the bounding volume. As the three components evolve over time, the position, size, and orientation of the initial oriented bounding volume change continuously, resulting in a dynamic oriented bounding geometry that can characterize the evolution trajectory of each component in real time.

[0095] In this embodiment of the invention, by employing the Jacobi iteration method to accurately solve the eigenvalues ​​and eigenvectors of the covariance matrix, combining orthogonal projection to achieve crosstalk-free feature decoupling, and using incremental calculation-based sliding time window extreme value extraction, precise alignment of rotation matrices, and continuous and smooth dynamic update of bounding volume parameters, the technical problems of residual crosstalk, mismatch between oriented bounding volume and feature distribution, poor real-time performance of dynamic tracking, and easy occurrence of sudden changes in traditional feature decoupling methods are overcome. Thus, the three types of fault components are completely independently represented, and a dynamic geometric model that is highly consistent with the discrete distribution of features is constructed.

[0096] In a preferred embodiment of the present invention, step 3 above may include:

[0097] Step 3.1: Extract the principal axis direction vectors of the dynamically oriented enclosing geometry corresponding to adjacent running cycles, calculate the angular offset of the principal axis direction vectors relative to the standard coordinate reference axis, and obtain the principal axis deflection attitude. Specifically, this includes: establishing the standard coordinate reference axis through the device's initial power-on calibration process; after the device is powered on for the first time and initialized, continuously collecting multi-dimensional time-series feature fusion sets for 10 normal running cycles, each running cycle lasting 30 seconds; performing principal component analysis on each normal running cycle to obtain 10 sets of principal component feature vectors; calculating the arithmetic mean of the corresponding principal axes in the 10 sets of feature vectors to obtain three initial standard principal axis direction vectors; verifying the stability of the initial standard principal axis direction vectors; calculating the angular offset of each set of feature vectors from the initial standard principal axis direction vectors; if all offsets are less than 0.05 radians, then the initial standard principal axis direction vector is determined as the final standard coordinate reference axis; if there are offsets greater than or equal to 0.05 radians, then the data of the corresponding abnormal cycle is removed, and the average value is recalculated until the stability requirements are met; extracting the three principal axis direction vectors of the dynamically oriented enclosing geometry corresponding to the last time step of the current running cycle. , , And the three principal axis direction vectors of the dynamically oriented bounding geometry corresponding to the last time step of the previous adjacent running cycle. , , All principal axis direction vectors are three-dimensional unit vectors with a magnitude of 1. Calculate the angular offset of each principal axis direction vector relative to its corresponding standard coordinate reference axis. The formula for calculating the angular offset is:

[0098] ;

[0099] In the formula, This represents the angular offset of the k-th principal axis, where k takes the value 1, 2, or 3. This is the direction vector of the k-th principal axis in the current running cycle; Let be the standard coordinate reference axis direction vector corresponding to the k-th principal axis. Since all vectors are unit vectors, calculate the arithmetic mean of the angular offsets of the three principal axes using the following formula:

[0100] ;

[0101] In the formula, This is the comprehensive angular offset of the main axis deflection attitude, expressed in radians, with a value ranging from 0 to... The principal axis deflection attitude is obtained.

[0102] Step 3.2: Based on the principal axis deflection attitude, solve for the common intersection region of the dynamically oriented bounding geometries corresponding to adjacent running cycles in the 3D feature space. Calculate the ratio of the volume of the common intersection region to the total volume of each dynamically oriented bounding geometries to obtain the spatial overlap volume ratio. Specifically, this includes: based on the principal axis deflection attitude, using the separation axis theorem to solve for the common intersection region of the dynamically oriented bounding geometries corresponding to two adjacent running cycles in the 3D feature space. The separation axis set contains 15 direction vectors, which are the six principal axis direction vectors of the two bounding bodies. , , , , , And the nine direction vectors obtained by cross products of the principal axes of the two enclosing volumes. , , , , , , , , For each separation axis, calculate the projection intervals of the two bounding volumes on that axis. The minimum value of the projection interval is the minimum value of the projections of all vertices of the bounding volume onto that axis, and the maximum value is the maximum value of the projections of all vertices of the bounding volume onto that axis. If the projection intervals on all 15 separation axes overlap, then the two bounding volumes have a common intersection region; if the projection intervals on any separation axis do not overlap, then the two bounding volumes do not intersect, and the volume of the common intersection region is 0. When a common intersection region is detected, calculate the intersection lines of the six surfaces of the current bounding volume with the six surfaces of the previous bounding volume in sequence. Each intersection line consists of two intersection points. Collect the endpoints of all intersection lines, remove duplicate points, and construct the vertex set of the intersecting polyhedron. Use the divergence theorem to calculate the volume of the intersecting polyhedron. The calculation formula is:

[0103] ;

[0104] In the formula, The volume of the common intersection region; The number of vertices of the intersecting polyhedron; Let be the three-dimensional coordinates of the i-th vertex of the intersecting polyhedron, with the vertices arranged clockwise according to the right-hand rule. , , , , , Calculate the volume of two dynamically oriented bounding geometries, where the bounding geometries are cuboids. The volume calculation formula is: ,in , , Given the lengths of the three principal axes, calculate the union volume of the two dynamically oriented bounding geometries using the following formula:

[0105] ;

[0106] In the formula, Let the volume be the union volume; The volume of the geometry is dynamically oriented for the current running cycle; To calculate the volume of the dynamically oriented geometry in the previous adjacent running cycle, the ratio of the volume of the common intersection region to the volume of the union is calculated using the following formula:

[0107] ;

[0108] In the formula, The value of the spatial overlap volume ratio is 0 to 1, which is the percentage of the spatial overlap volume.

[0109] Step 3.3: Perform a weighted fusion operation based on the cosine value of the angular offset of the main axis deflection attitude and the proportion of the spatial overlap volume to determine the spatiotemporal collaborative association weights. Specifically, this includes: performing a weighted fusion operation based on the cosine value of the angular offset of the main axis deflection attitude and the proportion of the spatial overlap volume to determine the spatiotemporal collaborative association weights. The weighting coefficients are obtained through training with historical fault data. 1000 sets of normal operation state samples and 1000 sets of fault state samples are selected, and the parameters are optimized with fault identification accuracy as the objective function. Finally, the weighting coefficient for the cosine value of the angular offset is determined to be 0.4, and the weighting coefficient for the proportion of the spatial overlap volume is determined to be 0.6. The calculation formula for the weighted fusion is as follows:

[0110] ;

[0111] In the formula, The spatiotemporal collaborative association weight has a value range of 0 to 1; This is the weighting coefficient for the cosine value of the angle offset, with a value of 0.4; The weighting coefficient for the spatial overlap volume ratio is 0.6. The cosine value of the angle offset reflects the continuity of the equipment state in the time dimension. The closer the cosine value is to 1, the smaller the spindle deflection and the more stable the equipment state. The spatial overlap volume ratio reflects the overlap of the equipment state in the spatial dimension. The closer the ratio is to 1, the higher the overlap of the feature distributions of the two cycles and the smaller the change in the equipment state. The weighted spatiotemporal collaborative association weight after weighted fusion comprehensively characterizes the degree of collaborative change of the equipment state in the spatiotemporal dimension. The closer the weight is to 1, the more stable the equipment state is. The closer the weight is to 0, the more abrupt the equipment state changes.

[0112] Step 3.4: Based on the current operating cycle state corresponding to the spatiotemporal collaborative association weight, extract the spatial coordinates of the overlapping centroids of the dynamically oriented enclosing geometry. Construct a multidimensional reference datum surface using a preset standard operating boundary. Calculate the perpendicular distance from the overlapping centroid spatial coordinates along the normal of each feature dimension to the multidimensional reference datum surface to obtain the multidimensional feature deviation. Specifically, the preset standard operating boundary is obtained through statistical analysis of normal operating data. Collect dynamic oriented enclosing geometry data for 100 consecutive normal operating cycles of the equipment. Extract the maximum value of the feature components in the three principal axis directions for each cycle. Calculate the 95th quantile of the maximum value in each principal axis direction as the upper limit boundary for normal operation in that principal axis direction. Construct three multidimensional reference datum surfaces perpendicular to the three standard coordinate reference axes, with each reference surface corresponding to the upper limit boundary for normal operation in one principal axis direction. The plane equations of the multidimensional reference datum are:

[0113] ;

[0114] in, Let be the coordinates of any point in space. Let be any point on the reference plane, and let its coordinates be the point after moving the origin of the standard coordinate system by 95% quantile distance along the k-th principal axis. Expand the plane equation into its general form. ,in , , Standard coordinate reference axis The three components, When two dynamically oriented bounding geometries have a common intersection region, the centroid coordinates of the common intersection region are extracted as the overlapping centroid coordinates. When the two bounding geometries do not intersect, the centroid coordinates of the dynamically oriented bounding geometries in the current running cycle are taken as the overlapping centroid coordinates. The overlapping centroid coordinates are the arithmetic mean of the coordinates of all vertices of the corresponding polyhedron, calculated using the following formula:

[0115] ;

[0116] In the formula, The coordinates of the overlapping centroids in space; To determine the number of vertices of the corresponding polyhedron, calculate the perpendicular distance from the overlapping centroid spatial coordinates along the normal of each feature dimension to the corresponding multidimensional reference plane. The formula for calculating the perpendicular distance from a point to a plane is:

[0117] ;

[0118] In the formula, The deviation of the k-th feature dimension; The coordinates of the overlapping centroids in space; , , , Let be the plane equation coefficients of the k-th multidimensional reference datum. Since the datum normal vector is a unit vector, calculate the weighted average of the deviations in the three characteristic dimensions. The weighting coefficients are the proportions of the eigenvalues ​​corresponding to each principal axis to the total eigenvalues. The calculation formula is:

[0119] ;

[0120] In the formula, The overall multidimensional feature deviation; , , The eigenvalues ​​corresponding to the three principal axes are used to obtain the overall multidimensional feature deviation.

[0121] In this embodiment of the invention, a stable spatial orientation reference is established by adopting a standardized equipment power-on calibration process. The intersection region and volume of the oriented bounding volume are accurately solved using the separation axis theorem. Weighting coefficients are determined based on historical data training to calculate the spatiotemporal collaborative correlation weights. Furthermore, a standard operating boundary is established using statistical quantiles, and the orthogonal projection distance is calculated to obtain the multidimensional feature deviation. Therefore, this overcomes the technical problems of traditional fault judgment methods, such as unstable references, large errors in calculating intersection volumes, insufficient consideration of spatiotemporal correlation, and inaccurate quantification of state deviation. Consequently, correlation weights and deviation indices that accurately reflect the spatiotemporal evolution characteristics of the equipment state are obtained.

[0122] In a preferred embodiment of the present invention, step 4 above may include:

[0123] Step 4.1: Based on the spatiotemporal collaborative correlation weights and multidimensional feature deviations, using the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component as matrix dimension indices, the spatiotemporal collaborative correlation weights are mapped to the off-diagonal intersection nodes of the matrix, and the multidimensional feature deviations are mapped to the diagonal master nodes of the matrix. Tensor dimension splicing is performed to construct a multidimensional fault feature correlation matrix. Specifically, the multidimensional fault feature correlation matrix is ​​a 3×3 two-dimensional matrix. The row and column indices of the matrix are arranged in a fixed order according to the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component. The index numbers are 1, 2, and 3 respectively. The matrix elements are stored using 32-bit single-precision floating-point numbers with a value range of 0 to 1. This data type ensures calculation accuracy while adapting to the limited computing resources and storage capacity of industrial edge devices. The multidimensional feature deviations corresponding to the three components are mapped to the diagonal master node of the matrix in index order, i.e., the position of the i-th row and i-th column of the matrix. The diagonal elements represent the degree to which the corresponding fault component deviates from the normal operating state. The spatiotemporal co-correlation weights between any two components are mapped to the corresponding off-diagonal intersection node of the matrix, i.e., the position of the i-th row and j-th column where i≠j. Since the spatiotemporal co-correlation between two components is symmetrical, the correlation weights between the hardware electrical drift component and the communication link signal-to-noise ratio attenuation component are simultaneously filled into the 1st row and 2nd column and the 2nd row and 1st column, the correlation weights between the hardware electrical drift component and the micro-environmental disturbance component are simultaneously filled into the 1st row and 3rd column and the 3rd row and 1st column, and the correlation weights between the communication link signal-to-noise ratio attenuation component and the micro-environmental disturbance component are simultaneously filled into the 2nd row and 3rd column and the 3rd row and 2nd column. The expression for the multidimensional fault feature correlation matrix is:

[0124] ;

[0125] In the formula, A multi-dimensional fault feature correlation matrix; The multidimensional feature deviation corresponding to the hardware electrical drift component; The multidimensional feature deviation corresponding to the signal-to-noise ratio attenuation component of the communication link; The multidimensional feature deviation corresponding to the microenvironmental disturbance component; The spatiotemporal co-correlation weight between hardware electrical drift components and communication link signal-to-noise ratio attenuation components; The spatiotemporal co-correlation weight between hardware electrical drift components and micro-environmental disturbance components; The spatiotemporal co-correlation weight between the signal-to-noise ratio attenuation component and the hardware electrical drift component in the communication link; The spatiotemporal co-correlation weight between the signal-to-noise ratio attenuation component of the communication link and the micro-environmental disturbance component; The spatiotemporal co-correlation weight between micro-environmental disturbance components and hardware electrical drift components; To determine the spatiotemporal co-correlation weights between the micro-environmental disturbance component and the communication link signal-to-noise ratio attenuation component, a tensor dimension concatenation operation is performed. The input is a one-dimensional feature array containing 6 elements, with the array elements ordered as follows: , , , , , Following a row-major storage method, the elements of the one-dimensional array are sequentially filled into the corresponding positions of the two-dimensional matrix, where... , , Each by , , After copying and filling the matrix, perform a validity check to verify that all elements are within the range of 0 to 1, and verify the symmetry of the matrix, i.e., for all i ≠ j, the following condition is met. After successful verification, the final multi-dimensional fault feature correlation matrix is ​​obtained.

[0126] Step 4.2 involves inputting the multi-dimensional fault feature correlation matrix into a pre-defined fault mode identification strategy. This maps matrix nodes to entity nodes in the state transition network topology graph, and maps off-diagonal elements of the matrix to state transition edges between entity nodes. Specifically, the pre-defined fault mode identification strategy is implemented based on a directed weighted state transition network topology graph, which is stored using an adjacency list data structure. The adjacency list contains four head nodes, corresponding to the initial steady-state node and three entity nodes respectively. Each head node points to a linked list, and each node in the linked list stores information about a state transition edge. The structure of a linked list node includes an edge identifier, an endpoint node identifier, a weight value, and a pointer to the next linked list node. The multi-dimensional fault feature correlation matrix is ​​then input into the fault mode identification strategy, performing a mapping transformation from matrix to topology graph. The three fault components corresponding to the matrix dimension indices are mapped to three entity nodes in the state transition network topology graph, with the unique identifiers of the three entity nodes being... , These correspond to the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component, respectively.

[0127] Each entity node's attributes include a node type identifier, component name encoding, and a corresponding multidimensional feature deviation value. The node type identifier is an integer, with values ​​1, 2, and 3 corresponding to the three types of fault components, respectively. The component name encoding is a 16-bit unsigned integer used for rapid identification of the fault component type. Off-diagonal intersection nodes of the matrix are mapped to directed state transition edges between entity nodes. The starting point of each state transition edge is the entity node corresponding to the row index, and the ending point is the entity node corresponding to the column index. The weight of each state transition edge is the spatiotemporal collaborative association weight value of the corresponding off-diagonal intersection nodes of the matrix. Each state transition edge's attributes include an edge identifier, a starting node identifier, an ending node identifier, and a weight value. The edge identifier is a globally unique integer. The weight value of the state transition edge reflects the coupling degree between two corresponding fault components; a higher weight value indicates a stronger mutual influence between the two components, and a greater likelihood of fault propagation between them. After all nodes and edges have been created, perform a topology graph integrity verification. Check if all three entity nodes have been created, check if the adjacency list contains 6 directed state transition edges, and verify if the weight value of each edge is consistent with the corresponding element in the matrix. After the verification is passed, the construction of the state transition network topology graph is completed.

[0128] Step 4.3: Starting from the initial steady-state node, perform a depth-first path search along the state transition edges, recording the activation sequence and cumulative transition weights of each entity node in the path to obtain the feature-coupled topological state sequence. Specifically, this includes setting the initial steady-state node. As the starting point for depth-first path search traversal, the initial steady-state node is a virtual node corresponding to the normal operating state of the device, with an attribute value of 0 and an initial state transition edge weight of 0 between it and the three physical nodes. A boolean access marker array of length 4 is created, with the array index corresponding to the node identifier. Initially, all elements are false, indicating that all nodes have not been visited. A path stack is created, with each element containing four fields: node identifier, node attribute value, incoming edge weight, and activation timestamp. The maximum depth of the stack is 6, corresponding to a maximum search depth of 5 plus the initial steady-state node. As the starting point for traversal, it will contain Elements of node information are pushed onto the path stack. The element with index 0 in the access marker array is set to true. A path list is created to store all valid paths found. Each path element contains three fields: node identifier sequence, activation time sequence, and cumulative transition weight value. During the search, the current node is removed from the top of the path stack each time, and all state transition edges connected to the current node are obtained. Edges whose endpoint nodes have not been visited are filtered out. The filtered edges are quickly sorted in descending order of weight value. If there are unvisited edges and the current path depth is less than the maximum search depth of 5, the edge with the highest weight is selected, and the endpoint node of the edge is pushed onto the path stack. The element with the corresponding endpoint node index in the access marker array is set to true. The timestamp of the endpoint node being pushed onto the path stack is recorded as the activation time sequence of the node. The activation time sequence is based on the start time of the current running cycle, in seconds, and is calculated by subtracting the start time of the running cycle from the current system time. The product of the weight of the state transition edge and the attribute value of the endpoint node is added to the cumulative transition weight value of the current path. The formula for calculating the cumulative transition weight value is:

[0129] ;

[0130] In the formula, This represents the cumulative transition weights for the current path. This represents the number of entity nodes contained in the current path. For the first in the path The node to the first The weights of the state transition edges of each node; For the first in the path For each node's attribute values, if the current node has no unvisited edges or the current path depth has reached the maximum search depth, the node identifier sequence, activation sequence, and cumulative transition weight value of the current path are saved to the path list. Then, the current node is popped from the path stack, the element corresponding to the current node's index in the access marker array is set to false, and the search continues to backtrack to the previous node. When the path stack is empty, the search process terminates, and the path list is traversed to find the path with the largest cumulative transition weight value as the main evolution path. The identifiers, activation sequences, and corresponding cumulative transition weight values ​​of all entity nodes in the main evolution path are arranged in chronological order of activation to obtain the feature-coupled topological state sequence. Each element of the feature-coupled topological state sequence is a structure containing three fields: node ID, activation time, and cumulative contribution value.

[0131] In this embodiment of the invention, a structured multi-dimensional fault feature association matrix stored in 32-bit single-precision floating-point numbers is used to simultaneously characterize the self-deviation state and mutual coupling relationship of fault components. Combined with a directed weighted state transition network topology graph stored in an adjacency list, the matrix-to-graph mapping and integrity verification are achieved. Furthermore, the techniques of depth-first path search, backtracking mechanism, and evolution path selection based on weight priority sorting are employed. Therefore, the technical problems of traditional fault determination methods, such as difficulty in quantifying the coupling strength between fault components, inability to completely track fault propagation paths and evolution sequences, difficulty in distinguishing the evolution directions of primary and secondary faults, and data structures that are not suitable for the computing resources of industrial edge devices, are overcome. As a result, a feature-coupled topology state sequence that can accurately reflect the coupling relationship, propagation path, and evolution sequence of fault components is obtained.

[0132] In a preferred embodiment of the present invention, step 5 above may include:

[0133] Step 5.1: Analyze the node transition feature vectors in the feature-coupled topology state sequence, calculate the cosine similarity between the node transition feature vectors and the attribute vectors of each historical fault node in the pre-set industrial equipment fault knowledge graph, and filter out candidate fault nodes with similarity higher than the preset matching threshold. Specifically, this includes: analyzing the feature-coupled topology state sequence, extracting the node identifier, activation time, and cumulative contribution value of each entity node in sequence according to the activation time, and constructing a three-dimensional node transition feature vector. The node transition feature vector is stored in a 32-bit single-precision floating-point array with a length of 3. The first element is the node type identifier, which is completely consistent with the defined node type identifier. The values ​​1, 2, and 3 correspond to the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component, respectively. The second element is the activation time, based on the start time of the current operating cycle, in seconds, with precision retained to three decimal places. The third element is the cumulative transition weight value, which is the cumulative contribution value of the node in the evolution path, ranging from 0 to 1. The pre-built industrial equipment fault knowledge graph is stored in a graph database and contains 1200 historical fault nodes. Each historical fault node is associated with a set of attribute vectors extracted from 50 similar historical fault cases. The dimension of each attribute vector is consistent with the node's transition feature vector, and the meaning of each element is completely corresponding. The attribute vector set of historical fault nodes is obtained by clustering analysis of the feature-coupled topological state sequences of historical fault cases. Each set corresponds to a typical fault evolution pattern and its normal fluctuation range.

[0134] A three-dimensional feature space ellipsoidal inclusion matching algorithm is used for preliminary screening of historical fault nodes. This algorithm quantifies the spatial inclusion relationship between the current feature vector and the historical fault mode by constructing a confidence ellipsoidal inclusion of historical fault samples. It overcomes the shortcomings of traditional cosine similarity, which only considers vector direction and ignores the discreteness and correlation of feature distribution. For the attribute vector set of each historical fault node, its three-dimensional covariance matrix is ​​calculated. The formula for calculating the covariance matrix is ​​as follows:

[0135] ;

[0136] In the formula, Let f be the 3×3 covariance matrix of the f-th historical fault node; This represents the number of attribute vectors corresponding to the historical fault node, with a value of 50. Let i be the i-th attribute vector in this set; Given the mean vector of all attribute vectors in this set, solve for the covariance matrix. eigenvalues and the corresponding unit eigenvector , , The three eigenvectors form the three principal axes of the ellipsoid, constructing a 95% confidence level ellipsoidal containment. The lengths of the three semi-axes of the ellipsoid are as follows:

[0137] , , ;

[0138] In the formula The critical value for a chi-square distribution with 3 degrees of freedom at a 0.95 confidence level is 7.815. This value represents the transition feature vector of the current node. The transformation formula for converting to the local coordinate system of the ellipsoid is as follows:

[0139] ;

[0140] In the formula For the eigenvector , , The rotation matrix is ​​used to calculate the Mahalanobis distance from the current eigenvector to the center of the ellipsoid. The formula for calculating the Mahalanobis distance is:

[0141] ;

[0142] In the formula, The Mahalanobis distance; , , Let these be the three coordinate components of the current eigenvector in the local coordinate system of the ellipsoid. If the current feature vector is located within the confidence ellipsoid of the historical fault node, then the historical fault node is added to the preliminary candidate node set; if If a historical fault node is excluded, the cosine similarity between the current node's transition feature vector and the mean vector of the historical fault node is calculated for each node in the preliminary candidate node set. The cosine similarity is obtained by dividing the dot product of the two vectors by the product of their magnitudes, and its value ranges from 0 to 1. The closer the value is to 1, the more similar the feature patterns of the two vectors are. The preset matching threshold is 0.75, which is obtained through training with 1000 sets of labeled fault samples. During the training process, the goal is to optimize the comprehensive index of fault identification accuracy and false alarm rate. The final threshold is determined to ensure a fault identification rate of over 95% while controlling the false alarm rate to below 3%. Historical fault nodes with a cosine similarity higher than the preset matching threshold are selected and arranged in descending order of cosine similarity. The top 3 are selected as the final candidate fault node set. If the number of nodes with a similarity higher than the threshold is less than 3, all nodes that meet the conditions are selected.

[0143] Step 5.2: Based on the candidate fault nodes, perform a reverse tracing traversal along the directed causal relationship edges of the pre-set industrial equipment fault knowledge graph to identify the underlying hardware entity nodes that trigger the current state transition and determine the fault root cause category. Specifically, the pre-set industrial equipment fault knowledge graph adopts a directed weighted graph structure, which includes three types of nodes: fault phenomenon nodes, fault root cause nodes, and hardware entity nodes, as well as two types of directed edges: causal relationship edges and component relationship edges. The knowledge graph is stored using an adjacency list and hash index. Each node corresponds to a unique integer identifier. The hash table enables a fast mapping from node identifier to node attributes. Fault phenomenon nodes correspond to observable abnormal states during equipment operation, such as increased voltage ripple or increased communication error rate. Fault root cause nodes correspond to the direct causes of abnormal states, such as capacitor aging or poor interface contact. Hardware entity nodes correspond to specific physical components in the equipment, such as main control power modules or industrial Ethernet interface chips. Causal relationship edges point from fault root cause nodes to fault phenomenon nodes, and the edge weight represents the statistical probability that the root cause leads to the corresponding phenomenon, with a value ranging from 0 to 1. Composition relationship edges point from hardware entity nodes to fault root cause nodes, and the edge weight represents the statistical probability that the hardware entity has the corresponding fault, with a value ranging from 0 to 1. The knowledge graph contains 32 hardware entity nodes, 86 fault root cause nodes, and 127 fault phenomenon nodes, as well as 521 causal relationship edges and 94 composition relationship edges.

[0144] Based on the set of candidate fault nodes, a depth-first reverse traversal is performed from each candidate fault node, moving in the opposite direction of the causal edges, i.e., traversing from the fault phenomenon towards the root cause. During the traversal, a boolean access marker array of length 245 is created, with the array index corresponding to the node identifier. Initially, all elements are false, indicating that no node has been visited. A path stack is created, with each element containing four fields: node identifier, node type, incoming edge weight, and current path cumulative weight. The maximum depth of the stack is 4, corresponding to the maximum traversal depth of 3 plus the initial candidate fault nodes. The current candidate fault node is pushed onto the path stack, marking it as visited, and the current path cumulative weight is initialized to 1. Each time, the current node is removed from the top of the path stack, and all incoming causal edges of the current node are obtained. Edges that have not been visited by the starting node are filtered out, and the filtered edges are quickly sorted in descending order of weight value. If there are unvisited edges and the current path depth is less than the maximum tracing depth of 3, select the edge with the largest weight, push the starting node of the edge onto the path stack, mark the starting node as visited, and multiply the current path's cumulative weight by the weight of the edge to update the new cumulative path weight. If the current node has no unvisited incoming causal edges or the current path depth reaches the maximum tracing depth, determine whether the current node is a hardware entity node. If so, save the node sequence and cumulative weight of the current path to the tracing path list; then pop the current node from the path stack, set the element at the current node's index in the visit flag array to false, backtrack to the previous node, and continue searching for other unvisited incoming causal edges. When the path stack is empty, terminate the tracing traversal of the candidate fault node.

[0145] All tracing paths corresponding to all candidate fault nodes are compared, and the tracing path with the largest cumulative weight is selected. If there are two or more paths with a cumulative weight difference of less than 0.05, the multidimensional feature deviation of the hardware entity nodes corresponding to the endpoints of these paths is compared, and the path with the largest multidimensional feature deviation is selected as the final tracing path. The endpoint of this path is the underlying hardware entity node. Based on the type of the underlying hardware entity nodes and the description of the associated root cause nodes, the root cause categories are determined. These categories are divided into four main types: hardware aging, mechanical damage, environmental interference, and communication anomalies. Hardware aging includes subcategories such as capacitor capacity decay, resistor value drift, and semiconductor device parameter drift; mechanical damage includes subcategories such as connector loosening, solder joint cracking, and structural component deformation; environmental interference includes subcategories such as electromagnetic interference, exceeding temperature and humidity limits, and vibration and shock; and communication anomalies include subcategories such as interface chip damage, communication protocol errors, and link interruption. After each fault is handled, the fault case is added to the knowledge graph, and the weights of the corresponding causal edges and constituent edges are updated. The weight update uses a moving average method, with the new weight being 0.9 times the original weight plus 0.1 times the statistical probability of the fault.

[0146] Step 5.3: Based on the numerical mapping relationship between the topological hierarchy depth and multidimensional feature deviation of candidate fault nodes in the pre-set industrial equipment fault knowledge graph, calculate the fault impact diffusion radius; combine the fault impact diffusion radius with a preset severity grading threshold range to determine the level, and obtain the fault location and grading results. Specifically, the nodes in the pre-set industrial equipment fault knowledge graph are divided according to the topological hierarchy. The bottom hardware entity nodes are layer 0, the fault root cause nodes directly connected to the hardware entity nodes through association edges are layer 1, the fault phenomenon nodes directly connected to the layer 1 fault root cause nodes through causal association edges are layer 2, and the fault phenomenon nodes directly connected to the layer 2 fault phenomenon nodes through causal association edges are... The upper-level fault phenomenon node is the 3rd layer, and so on. The basis for the topology hierarchy is the causal relationship of fault propagation. The larger the hierarchy number, the closer the fault phenomenon is to the system level and the wider the impact range. The topology hierarchy depth of the candidate fault node is the hierarchy number of the node in the knowledge graph, with values ​​of 0, 1, 2, and 3. A numerical mapping relationship between the topology hierarchy depth and the deviation of multidimensional features is established, and the fault impact diffusion radius is calculated. The proportional coefficient γ is adaptively adjusted according to the fault root cause category. The value of γ is 1.2 for hardware aging faults, 1.3 for mechanical damage faults, 0.9 for environmental interference faults, and 1.1 for communication anomaly faults. The formula for calculating the fault impact diffusion radius is:

[0147] ;

[0148] In the formula, The radius of the spread caused by the fault is dimensionless. This is the proportional coefficient corresponding to the root cause category of the fault; The topological hierarchy depth of the candidate faulty node; The deviation of the multidimensional feature corresponding to the candidate fault node ranges from 0 to 1. The physical meaning of the fault impact diffusion radius is the number of equipment functional modules that the fault may affect. R=0.5 indicates that the fault only affects a single hardware entity, R=1.0 indicates that the fault affects one functional module, R=1.5 indicates that the fault affects multiple functional modules, and R≥1.5 indicates that the fault may cause the entire equipment to shut down. The preset severity classification threshold range is determined by statistical analysis of the impact degree and processing priority of 2000 historical fault cases.

[0149] The severity of the fault is divided into four levels, when When a level 1 fault occurs, corresponding to a minor fault that does not affect normal equipment operation, the fault information should be recorded and addressed during the next routine maintenance. At this time, it is a level 2 fault corresponding to a general fault, and the equipment performance deteriorates, requiring technical personnel to be arranged for on-site inspection and handling within 24 hours; when At this time, it is a Level 3 fault corresponding to a severe fault, with partial malfunction of the equipment, requiring on-site repair by technicians within 4 hours; when At this time, a Level 4 fault corresponds to an emergency fault, indicating a risk of equipment downtime. An audible and visual alarm must be triggered immediately to notify equipment management and maintenance personnel. Simultaneously, the equipment shutdown protection program is automatically executed. The calculated fault impact radius is compared with a preset threshold range to determine the fault severity level. The fault root cause category, underlying hardware entity node information, fault impact radius, and fault severity level are integrated to obtain standardized fault location and classification results. The results are output in JSON format, containing seven fields: unique fault identifier, fault occurrence time, fault location, fault root cause description, impact scope assessment, severity level, and recommended handling measures. The fault location is specified down to the specific hardware entity number and installation location; the fault root cause description includes the specific type of fault and possible causes; the impact scope assessment describes the equipment functional modules and production processes that the fault may affect; and the recommended handling measures are generated based on the fault severity level and root cause category.

[0150] In this embodiment of the invention, candidate fault nodes are screened by using a dual screening mechanism of three-dimensional feature space ellipsoidal inclusion matching algorithm combined with cosine similarity. A depth-first reverse tracing traversal is performed using a directed industrial equipment fault knowledge graph containing multi-level nodes and weighted associated edges. Furthermore, technical means such as calculating the fault impact diffusion radius based on adaptive proportional coefficients and objectively classifying the faults are employed. Therefore, the technical problems of traditional fault judgment methods, such as only considering feature direction similarity, ignoring feature distribution characteristics leading to high false alarm rates, inability to locate specific hardware entities, difficulty in distinguishing the impact degree of different fault propagation paths, and reliance on human experience for fault severity classification leading to inconsistent standards, are overcome. This enables high-precision automatic root cause localization, quantitative impact range assessment, and objective classification of faults in industrial edge front-end equipment.

[0151] like Figure 2 As shown, embodiments of the present invention also provide a multi-dimensional fault determination system for front-end devices, including:

[0152] The acquisition module is used to acquire multi-dimensional raw monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity conditions during the current operating cycle; it performs time-series benchmark alignment and transient interference adaptive filtering on the multi-dimensional raw monitoring data streams, extracts the steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel, and obtains a multi-dimensional time-series feature fusion set;

[0153] The module is used to perform feature decoupling and cross-domain mapping processing on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components; using the hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamic oriented enclosing geometry representing the evolution trajectory of each component is constructed;

[0154] The calculation module is used to determine the spatiotemporal collaborative association weight by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and to calculate the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary to obtain the multidimensional feature deviation.

[0155] The mapping module is used to superimpose the spatiotemporal collaborative correlation weights and multidimensional feature deviations through tensor mapping to construct a multidimensional fault feature correlation matrix; the multidimensional fault feature correlation matrix is ​​input into a preset fault mode identification strategy, and a topological state traversal operation is performed to obtain a feature-coupled topological state sequence.

[0156] The judgment module is used to perform node similarity matching and causal chain inference between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph to obtain fault location and classification judgment results.

[0157] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0158] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0159] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0160] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for multi-dimensional fault determination of front-end devices, characterized in that, The method includes: Collect multi-dimensional raw monitoring data streams of industrial edge front-end devices deployed under strong electromagnetic interference and alternating temperature and humidity conditions during the current operating cycle; perform time-series benchmark alignment and transient interference adaptive filtering on the multi-dimensional raw monitoring data streams, extract the steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel, and obtain a multi-dimensional time-series feature fusion set; Feature decoupling and cross-domain mapping are performed on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components. Using the hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamic oriented enclosing geometry representing the evolution trajectory of each component is constructed. The spatiotemporal collaborative association weights are determined by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary is calculated to obtain the multidimensional feature deviation. The spatiotemporal collaborative correlation weights and multidimensional feature deviations are superimposed by tensor mapping to construct a multidimensional fault feature correlation matrix; the multidimensional fault feature correlation matrix is ​​input into a preset fault mode identification strategy, and a topological state traversal operation is performed to obtain a feature-coupled topological state sequence. By performing node similarity matching and causal chain deduction between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph, the fault location and classification results are obtained.

2. The multi-dimensional fault determination method for front-end devices according to claim 1, characterized in that, Collect multi-dimensional raw monitoring data streams from industrial edge front-end devices deployed under conditions of strong electromagnetic interference and fluctuating temperature and humidity during their current operating cycle, including: Real-time reading of voltage ripple sampling values ​​and current load fluctuation sampling values ​​of the main control power supply circuit to obtain the original sequence of hardware electrical telemetry; synchronous capture of bit error rate count, carrier signal-to-noise ratio fluctuation value and link handshake status identifier of the physical layer of industrial communication interface to obtain the original sequence of communication link status; The output sampling values ​​of the temperature sensing unit, humidity sensing unit and triaxial vibration accelerometer on the inner wall of the protective chassis are collected to obtain the original sequence of microenvironment sensing; the interrupt response delay record, memory stack allocation fragmentation rate and task scheduling blocking time of the embedded real-time kernel are extracted simultaneously to obtain the original sequence of kernel running log. The four types of raw sequences are injected with absolute timestamps generated by a unified hardware clock source, and cross-domain data frame encapsulation and cache alignment are performed to obtain multi-dimensional raw monitoring data streams.

3. The multi-dimensional fault determination method for front-end devices according to claim 2, characterized in that, The multi-dimensional raw monitoring data stream is subjected to time-series baseline alignment and transient interference adaptive filtering. Steady-state electrical characteristic baselines and nonlinear dynamic evolution trajectories of each monitoring channel are extracted to obtain a multi-dimensional time-series feature fusion set, including: Based on the absolute timestamps in the multi-dimensional original monitoring data stream, resampling and phase compensation are performed on the discrete sampling sequences of each channel to eliminate the timing offset caused by transmission link jitter and obtain the aligned data sequences of each channel. The aligned data sequences of each channel are input into an adaptive threshold sliding filter window. The filter window width and attenuation coefficient are dynamically adjusted according to the transient peak waveform characteristics of strong electromagnetic pulses and the thermal inertia characteristics of ambient temperature changes to filter out non-faulty high-frequency noise and step interference, thus obtaining a purified data sequence. Low-pass smoothing is performed on the purification data sequence to extract the steady-state electrical characteristic baseline of each monitoring channel within the preset normal operating range; the residual sequence between the purification data sequence and the steady-state electrical characteristic baseline is calculated. Gradient differencing and nonlinear envelope tracking are performed on the residual sequence to extract the nonlinear dynamic evolution trajectory that characterizes the equipment condition deterioration trend. The steady-state electrical characteristic baseline and the nonlinear dynamic evolution trajectory are concatenated and mapped by feature vectors according to the channel dimension to obtain a multi-dimensional time series feature fusion set.

4. The multi-dimensional fault determination method for front-end devices according to claim 3, characterized in that, Feature decoupling and cross-domain mapping are performed on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components and micro-environmental disturbance components; Using hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamically oriented enclosing geometry characterizing the evolution trajectory of each component is constructed, including: The multi-dimensional temporal feature fusion set is mapped to a three-dimensional feature orthogonal space. By constructing a multi-dimensional feature covariance matrix and solving the principal component eigenvectors, the principal axis distribution direction of the feature space is determined. Based on the orthogonal projection of the characteristic space of the spindle distribution direction and the channel crosstalk stripping process, the decoupled output independently represents the hardware electrical drift component that characterizes the slow shift of the main control power supply circuit parameters, the communication link signal-to-noise ratio attenuation component that characterizes the degradation of the underlying signal quality of the communication interface, and the micro-environmental disturbance component that characterizes the superposition of thermal and humidity changes and mechanical vibration inside and outside the chassis. Calculate the spatial discrete distribution boundary of the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component within the sliding time window. Use the time series extreme point set of the three components as spatial vertex constraints, and rotate and align the coordinate axes along the direction of the principal component feature vector to obtain the initial oriented bounding body that encloses the discrete distribution of features within the current running cycle. Based on the temporal evolution of the three components, the coordinates of the centroid, the proportion of the principal axis length, and the spatial rotation attitude of the initial oriented bounding volume are continuously updated to obtain the dynamic oriented bounding geometry.

5. The multi-dimensional fault determination method for front-end devices according to claim 4, characterized in that, The spatiotemporal collaborative association weights are determined by solving the principal axis deflection attitude and spatial overlap volume ratio of dynamically oriented enclosing geometries, and the orthogonal projection distance from the centroid of each dynamically oriented enclosing geometry to the preset standard running boundary is calculated to obtain the multidimensional feature deviation, including: Extract the principal axis direction vector of the dynamically oriented enclosing geometry corresponding to adjacent running cycles, calculate the angular offset of the principal axis direction vector relative to the standard coordinate reference axis, and obtain the principal axis deflection attitude; Based on the main axis deflection attitude, the common intersection region of the dynamically oriented enclosing geometry corresponding to adjacent running cycles in the three-dimensional feature space is solved, and the ratio of the volume of the common intersection region to the total volume of each dynamically oriented enclosing geometry is calculated to obtain the spatial overlap volume ratio. A weighted fusion calculation is performed based on the cosine value of the angular offset of the main axis deflection attitude and the proportion of the spatial overlapping volume to determine the spatiotemporal collaborative association weight; Based on the current operating cycle state corresponding to the spatiotemporal collaborative association weight, the spatial coordinates of the overlapping centroids of the dynamically oriented enclosing geometry are extracted. A multidimensional reference datum is constructed with a preset standard operating boundary. The vertical distance from the spatial coordinates of the overlapping centroids along the normal of each feature dimension to the multidimensional reference datum is calculated to obtain the multidimensional feature deviation.

6. The multi-dimensional fault determination method for front-end devices according to claim 5, characterized in that, By superimposing the spatiotemporal collaborative correlation weights and the multidimensional feature deviations through tensor mapping, a multidimensional fault feature correlation matrix is ​​constructed. The multi-dimensional fault feature correlation matrix is ​​input into a preset fault mode identification strategy, and a topological state traversal operation is performed to obtain a feature-coupled topological state sequence, including: Based on the spatiotemporal co-correlation weights and multidimensional feature deviations, the hardware electrical drift component, the communication link signal-to-noise ratio attenuation component, and the micro-environmental disturbance component are used as matrix dimension indices. The spatiotemporal co-correlation weights are mapped to the off-diagonal intersection nodes of the matrix, and the multidimensional feature deviations are mapped to the diagonal master nodes of the matrix. Tensor dimension splicing is performed to construct a multidimensional fault feature correlation matrix. Input the multi-dimensional fault feature association matrix into the preset fault mode identification strategy, map the matrix nodes to entity nodes in the state transition network topology graph, and map the off-diagonal elements of the matrix to state transition edges between entity nodes. Starting from the initial steady-state node, a depth-first path search is performed along the state transition edges. The activation sequence and cumulative transition weight of each entity node in the path are recorded to obtain the feature-coupled topological state sequence.

7. The multi-dimensional fault determination method for front-end devices according to claim 6, characterized in that, By performing node similarity matching and causal chain deduction between the feature-coupled topological state sequence and a pre-set industrial equipment fault knowledge graph, fault location and classification results are obtained, including: The node transition feature vectors in the feature-coupled topological state sequence are analyzed, and the cosine similarity between the node transition feature vectors and the attribute vectors of each historical fault node in the pre-set industrial equipment fault knowledge graph is calculated. Candidate fault nodes with similarity higher than the preset matching threshold are then selected. Based on candidate fault nodes, reverse tracing is performed along the directed causal relationship edges of the pre-set industrial equipment fault knowledge graph to identify the underlying hardware entity nodes that trigger the current state transition and determine the fault root cause category. Based on the numerical mapping relationship between the topological hierarchy depth and multidimensional feature deviation of candidate fault nodes in the pre-set industrial equipment fault knowledge graph, the fault impact diffusion radius is calculated; the fault impact diffusion radius is combined with the preset severity classification threshold range to determine the level, and the fault location and classification results are obtained.

8. A multi-dimensional fault determination system for front-end devices, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to collect multi-dimensional raw monitoring data streams from industrial edge front-end devices deployed under conditions of strong electromagnetic interference and alternating temperature and humidity during the current operating cycle. The multi-dimensional raw monitoring data stream is subjected to time-series benchmark alignment and transient interference adaptive filtering. The steady-state electrical characteristic baseline and nonlinear dynamic evolution trajectory of each monitoring channel are extracted to obtain a multi-dimensional time-series feature fusion set. The module is used to perform feature decoupling and cross-domain mapping processing on the multi-dimensional temporal feature fusion set to obtain hardware electrical drift components, communication link signal-to-noise ratio attenuation components and micro-environmental disturbance components. Using hardware electrical drift components, communication link signal-to-noise ratio attenuation components, and micro-environmental disturbance components as inputs, a dynamic oriented enclosing geometry characterizing the evolution trajectory of each component is constructed; The calculation module is used to determine the spatiotemporal collaborative association weight by solving the principal axis deflection attitude and spatial overlap volume ratio of the dynamically oriented enclosing geometry, and to calculate the orthogonal projection distance from the centroid of the overlap of each dynamically oriented enclosing geometry to the preset standard running boundary to obtain the multidimensional feature deviation. The mapping module is used to superimpose the spatiotemporal collaborative correlation weights and multidimensional feature deviations through tensor mapping to construct a multidimensional fault feature correlation matrix. Input the multi-dimensional fault feature correlation matrix into the preset fault mode identification strategy, perform topological state traversal operation, and obtain the feature-coupled topological state sequence. The judgment module is used to perform node similarity matching and causal chain inference between the feature-coupled topological state sequence and the pre-set industrial equipment fault knowledge graph to obtain fault location and classification judgment results.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images