An adaptive fault diagnosis deep learning method for non-stationary signals

Abstract

The application provides a kind of adaptive fault diagnosis deep learning method for non-stationary signal, it is related to the field of fault diagnosis, comprising: the non-stationary signal is cut into analysis window and input into deep learning model to obtain fault discrimination result, and the main label stream is constituted, the category change of adjacent window discrimination result is compared to generate migration event sequence, and the completely static period of continuous no change is identified;With the characteristics of the output of deep learning model, the category migration that occurs inevitably drifts with working condition is judged to determine abnormal static period, and the output is divided into drift state, static state and transition state according to the migration event sequence, and a differentiated output strategy is used;The discrimination of abnormal static period is executed, the contrast label stream is generated by reversing the direction of dependence, the overall loss of response of model and the mismatch of discrimination dependence direction are distinguished by comparing two label streams, forced alarm or automatic track switching processing is taken respectively, the label stream is reconstructed after track switching and the output is recovered by accumulating during the restart period.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, specifically to an adaptive deep learning method for fault diagnosis of non-stationary signals. Background Technology

[0002] Vibration signals generated by industrial equipment such as rotating machinery during operation are mostly non-stationary due to continuous changes in rotational speed and load, and the fault characteristic frequency also drifts accordingly; accurate fault diagnosis of these signals is the key to ensuring the safe operation of equipment.

[0003] In recent years, data-driven methods based on deep learning have been widely used in industrial monitoring. These methods typically perform time-frequency transformation on the acquired signals to extract features, and then input the feature vectors into classifiers such as support vector machines or neural networks. The classifier outputs a discrimination result based on the feature-class mapping relationship learned during the training phase. The discrimination results at each time step are generated independently and directly used for condition monitoring.

[0004] The shortcomings of existing technologies are as follows: when the signal distribution changes, the mapping relationship established during the training phase is difficult to maintain its effectiveness, and the reliability of the discrimination results decreases accordingly. More importantly, existing methods lack the ability to perceive their own output state, cannot promptly identify that the discrimination results have become invalid, and have difficulty distinguishing between the two situations of "the actual state of the equipment has changed" and "the model's discrimination ability has decreased", resulting in insufficient credibility of monitoring results under non-stationary operating conditions. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an adaptive deep learning method for fault diagnosis of non-stationary signals. This method solves the problems mentioned in the background, such as the lack of perception and evaluation capabilities of the device's own output state, the inability to identify and judge failures in a timely manner, the difficulty in distinguishing between changes in the actual state of the device and the decline in judgment capabilities, which lead to insufficient reliability of monitoring results under non-stationary operating conditions.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: an adaptive fault diagnosis deep learning method for non-stationary signals, comprising: S1, dividing the non-stationary signal into an analysis window, inputting it into a deep learning model to perform fault state discrimination to obtain discrete category labels to form a main label stream; S2, compare the category symbols of discrete category labels of adjacent analysis windows in the main label stream window by window to generate migration event indicators to form a migration event sequence, and count the sum of migration event indicators in a consecutive preset number of analysis windows. When the sum is zero, it is determined to be a completely static period and recorded in the completely static period record table. S3. Based on the physical inevitability criterion that the health deep learning model outputs will inevitably undergo category migration with the shift in working conditions, the completely still period is determined as an abnormal still period and written into the abnormal stillness marker. S4, based on the migration event sequence, the main tag stream is divided into drift state, static state and transition state. In the drift state, output is temporarily suspended and evidence is accumulated. In the static state, misjudgment suppression is determined. In the transition state, normal output is performed. S5, after confirming the abnormal static period by the misjudgment suppression determination, the discrimination dependency direction is reversed for each analysis window within the abnormal static period to generate a control label stream. The structural relationship between the control label stream and the main label stream is compared to distinguish between the overall loss of response of the model and the mismatch of the discrimination dependency direction. Forced alarm processing or automatic track switching processing is then performed. S6, after the automatic track switching process is completed, the main label stream is reconstructed starting from the first discrete category label, and the discrete category labels are accumulated in the restart fallback state. After accumulating a full preset number of analysis windows, the output is restored.

[0009] Preferably, S1 includes: S11, dividing the non-stationary signal into analysis windows of a fixed length of continuous sampling points, each analysis window containing a preset number of continuous sampling points, the starting sampling points of adjacent analysis windows being spaced apart by a preset number of sampling points, and overlapping sampling points between adjacent analysis windows; S12, inputting the continuous sampling points in each analysis window into the deep learning model, the output layer of the deep learning model outputting discrete category labels of the fault state category to which the analysis window belongs, the discrete category labels being taken from a set of fault state categories containing a preset number of fault state categories; S13, sequentially appending and saving each of the discrete category labels according to the time order of the analysis windows to form the main label stream.

[0010] Preferably, S2 includes: S21, reading the discrete category labels of adjacent analysis windows in the main label stream window by window in chronological order, and comparing the category symbols of the discrete category labels of adjacent analysis windows; S22, when the category symbols of the discrete category labels of adjacent analysis windows are different, the migration event indicator value takes a first preset value, and when the category symbols are the same, the migration event indicator value takes a second preset value, the migration event indicator value of the starting analysis window takes a second preset value, and the structural migration event is an event in which the category symbols of the discrete category labels of adjacent analysis windows change; S23, arranging each migration event indicator value in chronological order to form the migration event sequence, and adding a migration event indicator value for each new analysis window.

[0011] Preferably, S2 further includes: S24, sliding a preset number of consecutive analysis windows in the main label stream in a step-by-step manner, and counting the sum of migration event indicators within the window interval. When the sum is zero, it is determined that the preset number of consecutive analysis windows constitute a completely static period, and there are overlapping analysis windows between two adjacent sliding window intervals; S25, merging adjacent and overlapping completely static periods into a single completely static period. The starting position of the merged completely static period is the minimum value of the starting positions of each overlapping period, and the ending position is the maximum value of the ending positions of each overlapping period. The merged completely static period is recorded one by one into the completely static period record table according to the starting position and the ending position.

[0012] Preferably, S3 includes: S31, determining the completely still period as the abnormal still period based on the physical necessity criterion; S32, writing the abnormal still period into the completely still period, the abnormal still period being stored as an additional field of the corresponding entry in the completely still period record table, the value of the additional field being pending review, confirmation, or cancellation, and setting the value of the additional field to pending review when writing; S33, performing a review on the abnormal still period, the review result being confirmation or cancellation, and fixing the value of the additional field after the review is completed.

[0013] Preferably, S4 includes: S41, determining the output structure state of each analysis window according to the migration event sequence in a step-by-step manner of sliding one analysis window at a time, and recording it as the current state identifier; S42, when the current state identifier is a drift state, setting candidate count variables for each fault state category, accumulating the discrete category labels of the current analysis window corresponding to the candidate count variables, and when switching from the drift state to the transition state, taking the fault state category with the largest value of the candidate count variable as the drift state accumulated result category as the current output of the first analysis window for recovery output, and discarding the drift state accumulated result when switching from the drift state to the stationary state; S43, when the current state identifier is a stationary state, calculating the signal change rate of each analysis window during the abnormal stationary period, and counting the proportion of the number of analysis windows with a signal change rate greater than a preset first threshold to the total number of analysis windows during the abnormal stationary period, confirming the abnormal stationary period when the proportion is greater than a preset second threshold, and canceling the abnormal stationary label and reclassifying it as a transition state when the proportion is not greater than the preset second threshold.

[0014] Preferably, S5 includes: S51, after confirming the abnormal static period through misjudgment suppression determination, when the main discrimination is based on time-positive dependence, the original signal segments corresponding to each analysis window within the abnormal static period are arranged in reverse time order to form time-reverse signal segments, and the time-reverse signal segments are input into the deep learning model to perform feature discrimination and generate time-reverse discrimination labels; when the main discrimination is based on feature domain dependence, it is changed to organization discrimination based on another feature domain complementary to the feature domain to generate feature domain complementary discrimination labels, wherein the feature domain is the set of feature dimensions for performing feature discrimination on the organization signal; S52, the time-reverse discrimination labels or the feature domain complementary discrimination labels are appended and saved to form the control label stream, and the control labels are statistically analyzed. When the sum of migration event indicators in the signature stream during the abnormal static period is zero, the failure cause is determined to be that the model as a whole has lost its response; when it is greater than zero, the failure cause is determined to be that the discrimination dependency direction is mismatched. S5 also includes a handling decision, S53, when the failure cause is that the model as a whole has lost its response, generating alarm information including the identifier of the monitored device, the start and end positions of the abnormal static period, the failure cause determination result and the alarm generation time, and outputting it to the alarm display terminal, marking the output data corresponding to the abnormal static period as untrusted data; when the failure cause is that the discrimination dependency direction is mismatched, reversing the discrimination dependency direction of the deep learning model to the reverse direction, and clearing the main label stream, the migration event sequence and the complete static period record table.

[0015] Preferably, S6 includes clearing and restarting: rebuilding the main label stream starting from the first discrete category label after the track change, placing the main label stream in a restart fallback state, accumulating only discrete category labels, not performing output structure state determination, and not submitting discrete category labels, and after accumulating a full preset number of analysis windows, restoring the output structure state determination and output process.

[0016] (III) Beneficial Effects

[0017] This invention provides an adaptive deep learning method for fault diagnosis of non-stationary signals. It has the following advantages: 1. This invention generates a migration event sequence by comparing the discrimination results of adjacent windows window by window, and divides the output process into drift state, stationary state and transition state according to the distribution of migration events; adopts a differentiated output strategy in different states to match the model output with the signal change state; suspends output and accumulates evidence during the working condition drift period to avoid unreliable results being submitted directly; improves the credibility of monitoring output under non-stationary working conditions.

[0018] 2. This invention is based on the physical inevitability criterion that the output of a deep learning model will inevitably undergo category migration as the operating conditions change. It identifies continuous periods of complete stillness without migration as abnormal still periods. For abnormal still periods, it reverses the discrimination dependency direction to generate a control label stream. By comparing the structural relationship between the control label stream and the main label stream, it distinguishes between the overall loss of model response and the mismatch of discrimination dependency direction. It performs forced alarm or automatic track switching respectively. After track switching, the label stream is reconstructed and the output is cumulatively restored after a restart period. This enables timely identification and targeted recovery of discrimination failure. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is an overall flowchart of an adaptive fault diagnosis deep learning method for non-stationary signals as described in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: This invention provides an adaptive fault diagnosis deep learning method for non-stationary signals. In one embodiment, it includes: defining the following symbols and terms: t is the analysis window number, y is the discrete category label, T is the migration event indicator, r is the signal rate of change, and E is the signal energy. For the t-th analysis window, it is assumed to contain N sampling points, and the sampling value is... , ,…, Then the signal energy is defined as: The main label stream refers to the discrete category label sequence arranged in chronological order according to the analysis window. Represents the discrete category label of the t-th analysis window; This represents the discrete category label for the (t-1)th analysis window; This represents the migration event indicator for the t-th analysis window; This represents the rate of change of the signal in the t-th analysis window; , Let be the signal energy of the t-th analysis window. Let be the signal energy of the (t-1)th analysis window, where the signal energy is the sum of the squares of all sampling points within the analysis window; The monitored equipment is identified by the equipment number registered when it leaves the factory; the deep learning model is a trained deep neural network classification model, and its network structure and training method adopt conventional methods in this field; the deep learning model is a convolutional neural network, which sequentially includes a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, and the output layer uses the soft max activation function to output discrete category labels; Discriminant dependency direction refers to the signal organization method used by the deep learning model when organizing input signals to perform feature discrimination; inversion direction refers to the discriminant organization direction adopted after inverting the discriminant dependency direction; the comparison label stream refers to the label stream generated by inverting the discriminant dependency direction for each analysis window during the abnormal static period, arranged in chronological order; model inference loss of response refers to the state in which the deep learning model loses its effective discrimination ability due to decision boundary saturation, parameter collapse, or signal link degradation; discriminant dependency direction mismatch refers to the state in which the deep learning model still has discrimination ability, but the signal organization method on which the discrimination is based is inconsistent with the actual carrying method of the signal features; failure cause determination result refers to the determination conclusion of model inference loss of response or discriminant dependency direction mismatch after structural relationship comparison processing; normal output process refers to the action of submitting the discrete category label of the current analysis window as the discrimination output; The dynamic scaling rule refers to the rule that the number of continuous windows used in the complete stillness determination scales up and down with the rhythm of the previous tag flow migration. The dynamic scaling rule applies to the complete stillness determination. The preset number refers to the number of continuous analysis windows used in the complete stillness determination, and the preset number is 10. The preset number of times refers to the threshold of the number of migration events used in the drift state determination, and the preset number of times is 3. The first preset value and the second preset value are two values ​​of the migration event indicator, with the first preset value being 1 and the second preset value being 0.

[0022] Step 1 uses the non-stationary vibration signal of the monitored equipment collected by the vibration sensor as input; the vibration sensor is installed on the bearing seat or housing surface of the monitored equipment, and a sampling point is collected at fixed sampling intervals. The sampling points are arranged continuously in time sequence to form the original signal sequence; the non-stationary signal refers to the vibration signal whose fault characteristic frequency continuously drifts under the condition of continuous change of the speed and load of the monitored equipment. The original signal sequence is divided into multiple analysis windows, each containing 256 consecutive sampling points. The starting sampling points of adjacent analysis windows are spaced 128 sampling points apart, meaning that there is a 50% overlap in sampling points between adjacent analysis windows. An analysis window refers to a fixed-length continuous sampling point in the original signal sequence. The deep learning model uses 256 consecutive sampling points within a single analysis window as input to perform feature discrimination. The output layer outputs the discrete category label of the fault state category to which the analysis window belongs. The discrete category label is taken from the fault state category set, which includes four categories: normal state, rolling bearing inner ring fault state, rolling bearing outer ring fault state, and rolling element fault state. The discrete category label refers to the discrete symbol that identifies the fault state category. The discrete category labels are appended and saved sequentially according to the time order of the analysis window, forming a main label stream. The t-th symbol in the main label stream is denoted as... , The main label stream represents the discrete category label of the t-th analysis window. The main label stream grows in real time as the analysis window is generated window by window. The main label stream is delivered as output to step two for processing.

[0023] Step two takes the main label stream output from step one as input and reads adjacent labels from the main label stream window by window in chronological order, i.e., the discrete category labels of the t-th analysis window. Discrete category labels of the (t-1)th analysis window The system compares whether the category symbols of the two are the same; if the category symbols are different, a structural migration event occurs at the t-th analysis window, and the migration event indicator is... A value of 1 indicates that no structural migration event has occurred if the category symbols are the same. (Migration event indicator value) The value is 0, that is =1 holds true when y(t) ≠ y(t-1). =0 holds true when y(t)=y(t-1). The structural migration event refers to the event in which the discrete category labels of adjacent analysis windows change their category symbols. The second analysis window is used as the starting point for calculating the migration event indicator. The migration event indicator T(1) of the first analysis window is set to 0. T(1) is set to 0 only to handle the case where there is no predecessor label at the starting boundary, and does not affect the subsequent sliding statistics results. The migration event indicators are arranged in chronological order to form a migration event sequence. Each new analysis window adds a migration event indicator, and the migration event sequence is updated in real time. To determine if a migration event sequence is completely stationary: slide 10 consecutive analysis windows across the label stream in a step-by-step manner, and count the sum of migration event indicators within each window interval. If the sum is 0, then the 10 consecutive analysis windows constitute a completely stationary period. The window intervals of two adjacent slides overlap by 9 analysis windows. It should be noted that the half-sampling point overlap between adjacent analysis windows in step one refers to the overlap at the sampling point level, while the overlap of 9 analysis windows between two adjacent slides in this step refers to the overlap at the analysis window level. The two have different meanings. Adjacent and overlapping completely still periods are merged into a single completely still period. The starting position of the merged completely still period is the minimum of the starting positions of the overlapping periods, and the ending position is the maximum of the ending positions of the overlapping periods. The merged result is written into a completely still period record table, which records the starting and ending positions of each completely still period and updates in real time as the sliding decision progresses. The migration event sequence and the completely still period record table are then delivered to step three for processing.

[0024] The input for step three is the migration event sequence output from step two and the table of completely still periods. Each completely still period in the table of completely still periods is read one by one, and the stillness criterion is reversed for each completely still period. Under non-stationary signal conditions, the speed and load of the monitored equipment change continuously, and the fault characteristic frequency shifts continuously with the operating conditions. The output main label stream of the health deep learning model will inevitably undergo periodic category shifts with the operating conditions, that is, the main label stream will inevitably experience structural shift events in any continuous time period; this law is used as a criterion of physical inevitability. Conventional diagnostic understanding interprets a completely still label stream as a stable diagnostic result. This invention reverses this understanding: a completely still period refers to a period in which the sum of migration event indicators is 0 within 10 consecutive analysis windows. The occurrence of a completely still period in the main label stream violates the physical necessity criterion. Therefore, a completely still period is judged as an abnormal still period. The structural fact that the main label stream is completely still is reversed from "stable diagnostic result" to "failure symptom of model inference losing response". An abnormal stillness flag is written for a completely still period. This flag is stored as an additional field of the corresponding entry in the completely still period record table. The field value can be one of three: pending review, confirmed, or revoked. When writing, the field value is set to pending review. If the completely still period record table is empty, that is, there is no completely still period, then no abnormal stillness flag is generated, and migration event detection continues to be performed on subsequent newly added tags. The abnormal static marker is delivered to step four for processing. Step four performs a review on it, and the review result is either confirmation or revocation. After the review is completed, the field value is fixed.

[0025] The input for step four is the migration event sequence output from step two, the abnormal stationary marker output from step three, and the original signal sequence. Based on the migration event sequence, the main label stream is divided into three output structure states: drift state, stationary state, and transition state. The output structure state of each analysis window is determined in real time by sliding one analysis window at a time. After scanning one analysis window, the output structure state of the current analysis window is determined and recorded as the current state identifier, which takes the value of one of the three states: drift state, stationary state, or transition state. The drift state refers to a state in which the main label stream experiences at least 3 structural migration events within 10 consecutive analysis windows, representing a period of frequent switching of fault state categories; the stationary state refers to a state in which the sum of migration event indicators in the main label stream is 0 within 10 consecutive analysis windows and there are abnormal stationary markers in the corresponding period, representing a period in which the label output remains completely unchanged; the transitional state refers to a state in which the main label stream is neither a drift state nor a stationary state, including: 1 to 2 structural migration events occurring within 10 consecutive analysis windows and no abnormal stationary markers in the corresponding period, or the sum of migration event indicators being 0 within 10 consecutive analysis windows but no abnormal stationary markers in the corresponding period; the transitional state represents a period in which the label output remains basically stable and there are no suspected failures. Read the signal segment from the original signal sequence that corresponds to the current analysis window, and calculate the rate of change of the signal in the current analysis window based on the signal energy. , This indicates whether the original signal is still changing within the current window interval; The three output structure states correspond to three output strategies: the drift state corresponds to the strategy of temporarily suspending output and accumulating evidence, the static state corresponds to the strategy of suppressing false judgments, and the transition state corresponds to the normal output process. The switching directions between the three states include switching from the drift state to the transition state, switching from the drift state to the static state, switching from the static state to the transition state, and switching from the transition state to the drift state. Each switching direction is determined by the real-time judgment result of the output structure state of the current analysis window. The current state identifier and the signal change rate are delivered to the subsequent stages.

[0026] In step four, when the current state is determined to be a drift state, a candidate count variable is set for each fault state category in the fault state category set. The candidate count variable corresponds one-to-one with the fault state category, and the initial value is 0, which is stored in memory. After each analysis window completes the discrimination in the drift state, the candidate count variable corresponding to the fault state category to which the discrete category label of the current analysis window belongs is incremented by 1. During the drift state, the label output switches frequently. If the label of each frame is submitted directly, it will cause output jitter and misjudgment propagation. Therefore, during the drift state, the output of discrete category labels is temporarily suspended, and only evidence accumulation is performed. When the current state identifier switches from the drift state to the transition state, the fault state category with the largest candidate count variable value is determined as the drift state cumulative result category, which is used as the current output of the first analysis window of the recovery output, and all candidate count variables are cleared to zero; if the candidate count variables of multiple fault state categories have the same maximum value, the fault state category that first reaches the maximum value of the candidate count variable is determined as the drift state cumulative result category - the candidate count variable records the time when it reaches the current value each time it is incremented by 1, and the category that first reaches the maximum value is determined accordingly; When the current state identifier changes from the drift state to the stationary state, the accumulated results of the drift state are abandoned, all candidate count variables are cleared to zero, no discrete category labels are output, and control is transferred to the stationary state misjudgment suppression determination.

[0027] In step four, when the current state is identified as a static state, the completely static time period corresponding to the abnormal static marker is read, the signal change rate of each analysis window in the static state is extracted, and the number of analysis windows with a signal change rate greater than a preset first threshold is counted and recorded as the number of change windows; at the same time, the total number of analysis windows in the completely static time period is counted and recorded as the total number of windows. If the proportion of the number of changing windows to the total number of windows is greater than the preset second threshold, it is determined that "the signal is changing but the label is not moving". The completely still period is confirmed as an abnormal still period in which the model inference loses response. The abnormal stillness mark is changed from pending review to confirmation. The output of any discrete category label is stopped and the control is transferred to the discrimination dependency direction inversion self-verification in step five. If the proportion of the number of changing windows to the total number of windows is not greater than the preset second threshold, it is determined that both the signal and the label remain stable, indicating that the device is in a true stable and healthy working condition. The abnormal static label corresponding to the completely static period is removed, and the current state is changed from static state to transitional state. Discrete category labels are output according to the normal output process of transitional state. During the static state, drift state evidence accumulation is not performed. After the abnormal static label is removed, the discrete category label of the current analysis window is output according to the transitional state. The principle of misjudgment suppression is as follows: Under real stable and healthy operating conditions, the signal energy remains stable and the signal change rate remains at a low value; while when the model inference fails, the signal continues to change and the signal change rate remains at a high value; therefore, the proportion of the signal change rate is used to distinguish between normal stillness and abnormal stillness; in this embodiment, the first threshold is preset to 2% and the second threshold is preset to 50%, both of which are design settings; the review of the abnormal stillness mark is executed immediately after writing, and the result is fixed after the review is completed, and subsequent steps only read the abnormal stillness mark status after the review is completed; When the current state is determined to be a transitional state, the normal output process for the transitional state is executed, and the discrete category label of the current analysis window is submitted as the discrimination output.

[0028] In step five, after confirming the abnormal static period in the misjudgment suppression determination, the discrimination dependency direction is reversed, including two methods: time sequence reversal and feature domain reversal; the feature domain refers to the set of feature dimensions of the signal that are organized for feature discrimination. The principle of reversing the discriminant dependency direction is as follows: If the deep learning model loses its inference response, the model cannot produce effective discrimination regardless of the signal organization method of the input, and the output remains static after reversal; if only the discriminant dependency direction is mismatched, the actual signal feature carrying method remains unchanged, and reorganizing the signal after reversal can restore effective discrimination, and the output recovers the category transfer; therefore, reversing the discriminant dependency direction can be used to distinguish between two types of failure root causes. When the primary discrimination is based on positive time dependence, a time-series reversal is employed: the original signal segments corresponding to each analysis window and its 32 preceding analysis windows within the abnormal static period are rearranged in reverse time to form time-reversed signal segments. These segments are then input into the deep learning model for feature discrimination, generating time-reversed discrimination labels. If the number of preceding windows is less than 32, the time-reversed rearrangement is organized using the actual number of preceding windows. If the actual number of preceding windows is 0, only the original signal segments corresponding to the current analysis window are rearranged in reverse time. When the primary discrimination criterion is based on feature domain dependency, feature domain inversion is employed: the discrimination is organized using another feature domain that is complementary to the feature domain, with time-domain features and frequency-domain features being complementary to each other; when the input of the deep learning model contains only the single feature domain of time-domain waveform, feature domain inversion organizes the discrimination using the frequency-domain spectrum obtained from the discrete Fourier transform of the time-domain waveform, generating complementary feature domain discrimination labels; when the primary discrimination criterion is based on both time-positive dependency and feature domain dependency, time-series inversion is employed. The time-reverse discrimination labels or feature domain complementary discrimination labels are appended and saved sequentially according to the time order of the analysis window to form a control label stream, which is then delivered for structural relationship comparison processing.

[0029] In step five, the structural relationship comparison is performed after the control label stream is generated. The comparison is made on the structural relationship between the two label streams, namely whether the migration event appears in the control label stream and whether the main label stream and the control label stream are both static. The structural relationship comparison only compares the migration event distribution structure of the two types of label streams and does not compare any discrimination scores. The sum of migration event indicators in the main label stream during the abnormal static period is recorded as the first total migration. The first total migration is determined by the attribute that the abnormal static period itself is a completely static period, and its value is 0. The sum of migration event indicators in the control label stream during the abnormal static period is recorded according to the same migration event statistics rules as the main label stream, and is recorded as the second total migration. That is, the category symbols of discrete category labels in adjacent analysis windows in the control label stream are compared window by window. If the category symbols are different, a migration event is counted. If the total amount of the second transfer is 0, then the control label stream and the main label stream are completely static, indicating that no matter how the discrimination dependency direction is organized, the deep learning model cannot generate category transfer, and the failure reason is that the model as a whole has lost its response; if the total amount of the second transfer is greater than 0, then the control label stream shows category transfer, indicating that the deep learning model has recovered category transfer after changing the discrimination dependency direction, and the failure reason is that the discrimination dependency direction is mismatched; a single transfer event in the total amount of the second transfer constitutes category transfer; the failure reason is delivered to the disposal decision-making stage for processing.

[0030] In step five, the decision-making process executes either mandatory alarm processing or automatic track switching processing based on the failure cause determination result. Mandatory alarm processing and automatic track switching processing are mutually exclusive. Each abnormal static period is executed independently once. The failure cause determination result for the same abnormal static period is unique, and different abnormal static periods are determined independently. When the failure cause determination result is that the model as a whole has lost its response, a forced alarm process is executed. The forced alarm process is as follows: generate alarm information, which includes the identifier of the monitored device, the start position of the abnormal static period, the end position of the abnormal static period, the failure cause determination result, and the alarm generation time. The identifier of the monitored device is the device number registered when the monitored device leaves the factory. The alarm information is output to the alarm display terminal. All output data corresponding to the abnormal static period is marked as unreliable data. Unreliable data does not participate in subsequent diagnostic decisions. The normal output process of the monitored device is stopped. The normal output process refers to the action of submitting the discrete category label of the current analysis window as the discrimination output. When the failure cause is determined to be a mismatch in the discrimination dependency direction, an automatic track switching process is performed. The automatic track switching process is as follows: the discrimination dependency direction of the deep learning model is reversed to the reverse direction, that is, the deep learning model organizes the input signal in the reverse direction to perform feature discrimination, clears the legacy main label stream, migration event sequence, completely static period record table and candidate count variables stored in memory, the clearing scope is the above data stored in memory, and the main label stream is cleared and restarted.

[0031] Step Six: After the automatic track switching process completes the determination of the dependency direction reversal, execute the zeroing and restart process: Reconstruct the main label stream starting from the first discrete category label output after track switching. During the restart, place the main label stream in the restart catch-up state. In the restart catch-up state, only discrete category labels are accumulated, and the output structure state determination and discrete category labels are not performed or submitted. After the main label stream has accumulated discrete category labels for 10 consecutive analysis windows, the output structure state determination and output process are resumed. During the restart, the complete static determination is performed in a step-by-step manner of sliding one analysis window at a time. The backup state during the restart period is used to accumulate enough label samples again after the decision dependency direction is reversed, so as to avoid making incorrect judgments based on incomplete main label streams. The dynamic scaling rule applies to the complete static determination: the change in equipment speed is the ratio of the difference between the maximum and minimum speed values ​​within the signal period corresponding to 10 consecutive analysis windows to the rated speed; periods where the change in equipment speed is lower than the preset third threshold are determined to be short periods. Within a short period, the migration rhythm of the previous main tag stream is used as a reference. The number of consecutive windows used for the dynamic scaling complete static determination is as follows: when the previous main tag stream has 6 or more structural migration events within 10 consecutive analysis windows, the number of consecutive windows is reduced to 8; when 4 or fewer structural migration events occur, the number of consecutive windows is increased to 15; when 5 structural migration events occur, the number of consecutive windows remains at 10. In this embodiment, the preset third threshold is 5%, and the preset third threshold, the 6th threshold, and the 4th threshold are design settings. The dynamic scaling rule only applies to the complete static determination and does not apply to the backup state during the restart period.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive fault diagnosis deep learning method for non-stationary signals, characterized in that, include: S1, the non-stationary signal is divided into an analysis window, which is input into the deep learning model to perform fault state discrimination and obtain discrete category labels to form the main label stream; S2, compare the category symbols of discrete category labels of adjacent analysis windows in the main label stream window by window to generate migration event indicators to form a migration event sequence, and count the sum of migration event indicators in a consecutive preset number of analysis windows. When the sum is zero, it is determined to be a completely static period and recorded in the completely static period record table. S3. Based on the physical inevitability criterion that the health deep learning model outputs will inevitably undergo category migration with the shift in working conditions, the completely still period is determined as an abnormal still period and written into the abnormal stillness marker. S4, based on the migration event sequence, the main tag stream is divided into drift state, static state and transition state. In the drift state, output is temporarily suspended and evidence is accumulated. In the static state, misjudgment suppression is determined. In the transition state, normal output is performed. S5, after confirming the abnormal static period by the misjudgment suppression determination, the discrimination dependency direction is reversed for each analysis window within the abnormal static period to generate a control label stream. The structural relationship between the control label stream and the main label stream is compared to distinguish between the overall loss of response of the model and the mismatch of the discrimination dependency direction. Forced alarm processing or automatic track switching processing is then performed. S6, after the automatic track switching process is completed, the main label stream is reconstructed starting from the first discrete category label, and the discrete category labels are accumulated in the restart fallback state. After accumulating a full preset number of analysis windows, the output is restored.

2. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 1, characterized in that: S1 includes: S11, the non-stationary signal is divided into analysis windows according to continuous sampling points of a fixed length. Each analysis window contains a preset number of continuous sampling points. The starting sampling points of adjacent analysis windows are spaced apart by a preset number of sampling points, and there are overlapping sampling points between adjacent analysis windows. S12, the continuous sampling points in each analysis window are input into the deep learning model. The output layer of the deep learning model outputs discrete category labels of the fault state category to which the analysis window belongs. The discrete category labels are taken from a set of fault state categories containing a preset number of fault state categories. S13, each discrete category label is appended and saved sequentially according to the time order of the analysis window to form the main label stream.

3. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 2, characterized in that: S2 includes: S21, Read the discrete category labels of adjacent analysis windows in the main label stream window by window in chronological order, and compare the category symbols of the discrete category labels of adjacent analysis windows; S22, When the category symbols of the discrete category labels of adjacent analysis windows are different, the migration event indicator value takes a first preset value, and when the category symbols are the same, the migration event indicator value takes a second preset value. The migration event indicator value of the starting analysis window takes a second preset value. The structure migration event is an event in which the category symbols of the discrete category labels of adjacent analysis windows change. S23, arrange the migration event indicators in chronological order to form the migration event sequence, and add a migration event indicator for each new analysis window.

4. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 3, characterized in that: S2 further includes: S24, sliding a preset number of consecutive analysis windows in the main label stream in a step-by-step manner, and counting the sum of migration event indicators within the window interval. When the sum is zero, it is determined that the preset number of consecutive analysis windows constitute a completely static period, and there are overlapping analysis windows between two adjacent sliding window intervals; S25, merging adjacent and overlapping completely static periods into a single completely static period. The starting position of the merged completely static period is the minimum value of the starting positions of the overlapping periods, and the ending position is the maximum value of the ending positions of the overlapping periods. The merged completely static period is recorded one by one into the completely static period record table according to the starting position and the ending position.

5. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 4, characterized in that: S3 includes: S31, the completely still period is determined as the abnormal still period based on the physical necessity criterion; S32, Write the abnormal stillness marker to the completely still period. The abnormal stillness marker is stored as an additional field of the corresponding entry in the completely still period record table. The value of the additional field is pending review, confirmation, or cancellation. When writing, the value of the additional field is set to pending review. S33, Perform a review on the abnormal stillness marker. The review result is confirmation or cancellation. After the review is completed, the value of the additional field is fixed.

6. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 5, characterized in that: S4 includes: S41, determine the output structure state of each analysis window according to the migration event sequence by sliding one analysis window at a time, and record it as the current state identifier; S42, when the current state is identified as a drift state, candidate count variables are set for each fault state category. The discrete category labels of the current analysis window are accumulated according to the candidate count variables. When switching from the drift state to the transition state, the fault state category with the largest value of the candidate count variables is taken as the drift state cumulative result category and used as the current output of the first analysis window for recovery output. When switching from the drift state to the stationary state, the drift state cumulative result is discarded. S43, when the current state is identified as a stationary state, the signal change rate of each analysis window during the abnormal stationary period is calculated. The proportion of the number of analysis windows with a signal change rate greater than a preset first threshold to the total number of analysis windows during the abnormal stationary period is counted. When the proportion is greater than a preset second threshold, the abnormal stationary period is confirmed to be established. When it is not greater than the preset second threshold, the abnormal stationary label is removed and the state is changed to a transition state.

7. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 6, characterized in that: S5 includes: S51, after confirming the abnormal static period through misjudgment suppression, when the main discrimination is based on time-dependent correlation, the original signal segments corresponding to each analysis window within the abnormal static period are arranged in reverse time to form time-reverse signal segments. These time-reverse signal segments are then input into the deep learning model to perform feature discrimination, generating time-reverse discrimination labels. When the main discrimination is based on feature domain dependence, it is changed to organization discrimination using another feature domain complementary to the feature domain, generating feature domain complementary discrimination labels. The feature domain is the set of feature dimensions for performing feature discrimination on the organization signal. S52, the time-reverse discrimination labels or the feature domain complementary discrimination labels are appended and saved to form the control label stream. The control label stream is statistically analyzed within the specified range. The sum of migration event indicators within the abnormal static period is zero. When the sum is zero, the failure cause is determined to be that the model as a whole has lost its response. When the sum is greater than zero, the failure cause is determined to be that the discrimination dependency direction is mismatched. S5 also includes a handling decision. S53: When the failure cause is that the model as a whole has lost its response, an alarm message is generated containing the identifier of the monitored device, the start and end positions of the abnormal static period, the failure cause determination result and the alarm generation time. The alarm message is output to the alarm display terminal. The output data corresponding to the abnormal static period is marked as untrusted data. When the failure cause is that the discrimination dependency direction is mismatched, the discrimination dependency direction of the deep learning model is reversed to the reverse direction. The main label stream, the migration event sequence and the complete static period record table are cleared.

8. The adaptive fault diagnosis deep learning method for non-stationary signals according to claim 7, characterized in that: S6 includes zeroing and restarting: the main label stream is rebuilt starting from the first discrete category label after the track change, the main label stream is placed in the restart period fallback state, only discrete category labels are accumulated, the output structure state determination is not performed, and discrete category labels are not submitted. After accumulating a full preset number of analysis windows, the output structure state determination and output process are restored.

Images