The invention relates to a method of continuously monitoring the operational state of a machine, particularly a complicated machine with at least two rotating machine parts, working at different revolutionary rates and in mutual co-action, sensed vibration states being processed by analysis of frequency spectra while utilizing sampling and pattern recognition techniques, and abnormal operational conditions being detected by calculating the probability of a sensed vibration state differing significantly from normal operational statβB, which are represented by a reference class calculated on the baβis of previously sensed vibration states during normal operation of the machine .

Such a method is already known from EP-A-84902732.1 , where the reference class comprises frequency spectra and where pattern recognition and detection means are adapted for calculating the probability for each new frequency spectrum that the latter is associated with a class other than the reference class, whereby an abnormal operational state of the machine is detected when this probability exceeds a predetermined limit.

The known method is advantageous in that no interpretation of frequency spectra needs to be made as long as each sensed frequency sprectrum belongs to the reference class. Only when abnormal operational states occur does the frequency spectrum need to be studied more closely. For simple machines, each peak in the spectrum can be attributed to a given function or to a given machine part, and even very small functional changes can thus be discovered at an early stage.

In more complicated machines with at least two rotating machine parts operating at different revolutionary rates and in mutual co-action, particularly via different mechanisms, as is

the case in jet engines, the known method cannot be used without complications. Accordingly, each spectrum peak must be analysed with relation to its origin, which is complicated, and in addition iβ not always possible, since different machine parts in certain combinations of revolutionary rates can give rise to coinciding spectrum peaks.

Against this background the object of the present invention is to develop the known method such that monitoring will also be reliable, and the diagnosis of functional disturbances will be enabled for complicated machines of the kind indicated above.

This object is achieved in accordance with the invention by the measures disclosed in the characterizing portion of claim 1. Accordingly, the principle of directly applying pattern sensing techniques to sensed vibration spectra and their peaks is abandoned. Instead, a mutual adjustment of the peaks in sensed and theoretically calculated spectra is carried out for the purpose of forming so-called weighting vectors, the components of which are directly assignable to different machine parts or partial systems in the machine. Accordingly, a transformation from vibration spectra to such weighting vectors takes place before a statistic model is constructed and comparison between new and earlier states takes place.

Advantageous applications of the method in accordance with the invention are disclosed in claims 2 - 6. In detecting an abnormal operational state, a fault diagnosis can be made in a simple way, as disclosed in claim 6, since deviating components in the weighting vector can be directly identified and related to specific machine parts or partial systems in the machine.

The invention will now be described in more detail, and with reference to the accompanying drawings, which illustrate a preferred embodiment.

Fig. 1 very schematically illustrates a measuring syβtem with associated computer equipment for using the method in accordance with the invention, and

Fig. 2 is a block diagram of the essential steps in the method in accordance with the invention.

In Fig. 1 there are thus illustrated, much simplified, a plurality of vibration sensing βensorβ sj, 83, ..., β _n , which are disposed on different parts of an unillustrated machine, and in this case the machine is assumed to include two rotat¬ ing machine parts (shafts) operating at mutually different revolutionary rates n^ and n2. In addition to the βensprs 8 _j _, ^s 2» • • • » ^B n the measuring system also includes two transducers- for measuring the rates n^ and n2.

As described in more detail in the above-mentioned EP-A-Θ4902732.1 , the vibration sensors are each coupled to an amplifier a-j_, &2 , . . - , a _n » which in turn iβ connected, posβibly via an unillustrated filter, to a separate input on an A/D converter 1, forming together with the amplifiers a sampling means 2. The signals from the sensors sj, B2. ... , s _n are sampled under the control of a microprocessor 3, which is also directly connected to the transducers for the revolutionary rates n^ and 112. the signals also being amplified and digitized to form time series, which are transmitted together with the revolutionary rate signals to a monitoring computer 4, e.g. a personal computer, for further processing and analysis. The computer 4 and microprocessor 3 are mututally connected for data transmission and control in both directions in the way described in the above-mentioned EP publication, possibly via a remote communication link. In a special application of the invention the machine comprises an aircraft jet engine, however, the sampling means 2 and microprocessor 3 then being placed close to the engine, while the monitoring computer 4 is centrally

placed in the aircraft cockpit. Alternatively, the jet engine can be ground-tested, when ground tests are being performed on the engine, the computer equipment then being placed outside the aircraft .

Signal processing is carried out in accordance with the block diagram of Fig. 2. After sampling the vibration signals from the sensors B , 8 , .... β _n . each time series transmitted to the computer 4 is converted by a Fourier transform (FFT) into a frequency spectrum in the form of a table with levels and frequencies). A predetermined number M of the higheβt peaks are selected in this frequency spectrum.

In accordance with the invention, these sensed spectrum peaks. are compared with pre-calculated theoretical peaks associated with the respective machine part or partial system in the machine. During this calculation it is assumed that each machine part E^, with the intermediary of the respective mechanism M _j , generates a plurality of spectrum peaks ^ _j ^ with frequencies ^i k' the latter being dependent of the revolutionary rates. The sub-index k refers here to the respective harmonic. The revolutionary rates 'nj and n2 sensed by the transducers in the particular case are used in the calculation. The total number <N) of peaks in the theoretically calculated vibration spectrum is thus :

N = Σ Σ Σ N _ijk i j k

In certain combinations of revolutionary rates, it can occur that two or more of the theoretically calculated peaks are at the same frequency, but this relationship is accidental and disappears when the revolutionary rate relationship changes.

Each of the M selected peaks in a sensed, actual frequency spectrum is compared with the theoretically calculated peaks in

the appropriate frequency range associated with the respective machine part. For each machine part E^ the true and theoretical- ly calculated peaks are matched with each other, i.e. each actual peak is assigned one or more adjacent, theoretically calculated peaks. For each such match or assignation, the computer calculates an adjustment weight w^ _jjς , which is propor¬ tional to the height of the actual peak above the background level and is inversely proportional to the frequency distance between both peaks (the actual and the theoretically calculat- ed) .

For the machine part E^, under discussion, the different adjustment weights ^j^ are summed to form a part weight (the total weights for the part) associated with the respective machine part, as follows:

i = Σ Σ w _ijk 3 k

the process is then repeated for remaining machine parts E^ and their associated part weights ^ are formed, which together form a weight vector associated with the machine in its entirety, :

(Wi _n )

The components of which constitute a measure of the respective machine part contribution to the vibration spectrum.

For describing different parts of the machine or its different functions, e.g. phenomena related to revolutionary rate or gear tooth mesh, measurements are sometimes required within differen frequency ranges. The part weights built up from spectra within different frequency ranges can be combined while taking into account the resolution in the respective spectrum. The high-resolution spectra shall here be given greater weight, e.g.

the part weights can be summed after multiplication, each with a factor 1/B, where B is the bandwidth corresponding to the resolution in the respective spectrum.

The weight vectors calculated are used in the same way as the vibration spectra in the method according to the above-mentioned EP publication. Accordingly, a βpecial pattern recognition program (SIMCA or the like specially adapted program) is used for forming a statistic model of the normal machine function, namely in the form of a reference class. During continuous monitoring of the machine each new weight vector is compared (one for each spectrum or group of spectra within different fequency ranges) with the reference class. The distance from the reference class, expressed in a statistical spread value, decides whether the operational state under consideration differs significantly from the normal state.

In thiβ way, abnormal operational states can be detected with great reliability, even for comparatively minor functional disturbances. Since the components (the part weights) of the weight vectors have a direct relationship with specific machine parts, a fault diagnosis can easily be made.

The method in accordance with the invention can of course be applied to comparatively simple machines, e.g. these with only one basic revolutionary rate. In such applications also, there is achieved greater reliability and simplier diagnosis of possible operational disturbances.